JP2003175480A - Robot device and method of controlling behavior thereof, and associative memory and associative storage method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To apply a competitive neural network to conduct learning allowing additional learning and resistant against a noise under an actual environment. <P>SOLUTION: The competitive neural network is applied as a storage model for this robot device 1 to realize all the long-term memory by the same technical model, and the additional learning is thereby realized in the long-term memory. Statistical learning is carried out to realize the long-term memory less affected by the noise. An associative type meaning is stored to associate a meaning of a body based on an imperfect data. An associative episode is stored to call up a place and a time in memory. <P>COPYRIGHT: (C)2003,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an intelligent robot apparatus and a method for controlling the behavior of an intelligent robot which controls a behavior according to a recognition result of an external stimulus or a change in an internal state. TECHNICAL FIELD The present invention relates to a robot apparatus for associative memory, an action control method thereof, an associative memory apparatus and an associative memory method.

More specifically, the present invention is a neural network
The present invention relates to a robot device that associatively stores a change in internal state from an external stimulus using a network and retains information obtained by learning such as the name of an object for a long time, and a behavior control method thereof.
In particular, the present invention relates to a robot device capable of additional learning and capable of strong learning against noise in a real environment, a behavior control method thereof, an associative memory device and an associative memory method.

[0003]

2. Description of the Related Art A neural network was originally a brain science term meaning an ecological neural network, but in recent years, it has come to be used to represent some kind of engineering calculation method. There is.

A neural network is a simplified model of a neural network in the human brain, and means a network in which nerve cell neurons are connected via synapses that pass signals in only one direction. Signal transmission between neurons is performed via synapses, and various information processing models are expressed by appropriately adjusting the resistance, that is, the weight of synapses.

According to control / information processing using a neural network as an engineering model, it is possible to directly deal with non-linear problems such as friction and viscosity, and a learning function is provided, so that it is possible to change parameter settings. There is an advantage that it becomes unnecessary. In addition, by imitating the structure of the human brain, the present invention aims to efficiently perform processing such as pattern recognition and associative memory that humans are good at.

Each neuron inputs outputs from one or more other neurons after weighting them with synapses,
When the sum of the input values exceeds a certain threshold value, "fire", that is, output is issued. In addition, it is possible to realize predetermined information processing by forming a myriad of connection relationships among neurons and adding weights to the respective connections, and by changing these weights, learning by neurons can be performed.

One of the information processings by the neural network is associative memory. The associative memory referred to here is a so-called self-associative associative memory, in which a plurality of patterns are stored in advance as a memory pattern and a pattern similar to a certain pattern among them is recalled. To tell.

[0008] For example, a robot device that performs an autonomous operation in response to a change in external stimulus or an internal state, and other high-level interactive systems, can store the name of an object by associatively storing the change in the internal state from the external stimulus. The information obtained by learning can be retained for a long period of time. Further, even when only an external stimulus consisting of incomplete data is given, the meaning of a certain object can be evoked by firing the corresponding neuron. In other words, by applying associative memory, it becomes possible to realize realistic communication with humans at a higher intellectual level.

As models of associative memory, for example, "Hopfield" and "Associatron (As
inter-connected neural networks such as "sociatron" are widely known. In all of these, in a neural network configuration in which neurons are connected to each other, one memory item is expressed and stored as a code representing the states of a plurality of neurons. In contrast to the conventional computer that addresses and stores all items, such an associative memory system is a memory system that does not have an address, and is considered to be closer to the human memory mechanism.

However, the associative memory mechanism based on these mutual coupling type neural networks generally has a problem that the memory capacity is small. The basic mechanism of associative memory is to store a certain input pattern by changing the weight of the connection between neurons.
For example, in the case of performing additional learning, if a certain input pattern is stored and the next pattern is further stored, there is a problem of "forgetting" that the previously stored pattern is lost.

Further, it is common to apply different models for each type of memory. For this reason, in order to implement a long-term memory mechanism in a single robot device or a single dialogue system, it is necessary to configure with multiple associative memory models, which makes it complicated and increases the cost accordingly. I will end up.

Further, in the case of an autonomous legged mobile robot or the like, since a relatively large work space is searched for in a routeless manner, the actual environment for memory / learning changes dynamically. For this reason, noise is greatly affected by changes in the actual environment, and associative memory is required to be robust against noise.

[0013]

SUMMARY OF THE INVENTION An object of the present invention is to provide an excellent robot apparatus capable of suitably associatively storing a change in internal state from an external stimulus, an action control method therefor, an associative memory apparatus and an associative memory method. To provide.

A further object of the present invention is an excellent robot capable of associatively storing a change in an internal state from an external stimulus by using a neural network and holding information obtained by learning such as a name of an object for a long period of time. An object of the present invention is to provide a device, a behavior control method thereof, an associative memory device, and an associative memory method.

A further object of the present invention is to provide an excellent robot apparatus capable of performing additional learning and capable of strong learning against noise in a real environment, an action control method thereof, an associative memory apparatus and an associative memory method. To do.

[0016]

The present invention has been made in consideration of the above problems, and a first aspect thereof is a drive unit for driving a joint or driving a dialogue with a user.
An associative storage unit that stores a predetermined input pattern and recalls a complete storage pattern from a missing input pattern, and an action control unit that controls the body operation by the drive unit based on the storage pattern recalled by the associative storage unit And a robot apparatus.

A second aspect of the present invention is a behavior control method for a robot apparatus for driving a joint or driving a dialogue with a user, which stores a predetermined input pattern and complete storage from a missing input pattern. An action control method for a robot device, comprising: an associative memory step of recalling a pattern; and an action control step of controlling a body motion of the robot device based on the memory pattern recalled by the associative memory step. Is.

Here, the associative storage unit or step is
A competitive neural network having an input layer consisting of a plurality of input neurons firing according to an input pattern, and a competitive layer consisting of a plurality of competitive neurons receiving weighted inputs of outputs from the respective input neurons Use associative memory.

That is, the associative memory unit or step selects a competitive neuron having a maximum sum of weighted input values from each input neuron for a certain input pattern, and selects the competitive neuron and each input neuron. By strengthening the binding force, the memory of the input pattern is strengthened.

Further, the associative memory unit or step selects a competitive neuron having a maximum sum of weighted input values from each input neuron with respect to an input pattern in which a part is missing, and selects a competitive force with the competitive neuron. You can recall a memory pattern consisting of a strong input neuron.

Therefore, according to the robot apparatus or the behavior control method thereof according to the first or second aspect of the present invention, by applying the competitive neural network as the memory model of the robot apparatus 1, all the long-term memory is achieved. Can be realized with the same engineering model.

Further, when a new pattern is input and an attempt is made to memorize, a new competitive neuron is fired and the connection with the new neuron is strengthened. However, since the connection of the neuron due to the previous memory is not weakened, the addition in the long-term memory is added. Learning can be realized.

Further, since the statistical learning is performed such that the memory is strengthened as the same pattern is input many times, it is possible to realize long-term memory with less influence of noise.

The robot apparatus or the behavior control method thereof according to the first or second aspect of the present invention may further include a sensory input unit for inputting an external stimulus.

In such a case, the associative storage unit or step can store the sensory input pattern from the sensory input unit and recall a complete sensory memory pattern from the sensory input pattern that is partially lost.

The associative storage unit or step stores an input pattern consisting of sensory inputs and internal states, and recalls a complete sensory memory pattern and / or internal state pattern from a sensory input pattern partially missing. You can That is, it is possible to associate the meaning of a certain object from the imperfect sensory input of the robot device.

The associative memory unit or step stores an input pattern consisting of a sensory input and an episode at the time of storage, and a complete sensory memory pattern and / or episode
Can recall patterns. That is, by performing the associative episode memory, the robot apparatus can associate an episode such as a place or a time at which a certain event is stored with an incomplete sensory input.

The associative memory unit or step stores an input pattern consisting of sensory input and motion and recalls a complete sensory memory pattern and / or a motion pattern of the drive unit from a sensory input pattern partially missing. Then, the operation of the machine can be controlled in accordance with the recalled operation pattern of the drive unit. That is, by performing the associative procedural knowledge storage, the robot apparatus can associate the stored procedure with the imperfect sensory input, that is, execute the operation. For example, if you look at something red and round, you can think of an apple and try to eat it.

In the competitive neural network applied in the associative memory unit or step, the competitive layer corresponds to each noun and a noun layer composed of a plurality of noun neurons storing nouns corresponding to input patterns. It may be configured with a verb layer composed of a plurality of verb neurons that store the verb to be stored.

In such a case, the noun neuron that first competes in a noun layer for a certain input pattern and wins is used to acquire a noun, and the acquired noun neuron competes in the verb layer to obtain the most noun neuron. Verb neurons with strong connections fire.

As described above, the efficiency of the calculation can be improved by dividing the calculation into the noun layer and the verb layer.
Further, when the number of verbs increases, it is necessary to tune between the noun neuron and the verb neuron when it is composed of a single competitive layer, but when the two layers are independent, this kind of complication occurs. Is released.

Further objects, features and advantages of the present invention are as follows.
It will be apparent from the embodiments of the present invention described later and the more detailed description based on the accompanying drawings.

[0033]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described below in detail with reference to the drawings.

A. Configuration of Robot Device FIG. 1 schematically shows a functional configuration of a robot device 1 used for implementing the present invention. As shown in FIG. 1, the robot apparatus 1 includes a control unit 20 that performs overall control of the entire operation and other data processing, an input / output unit 40, a drive unit 50, and a power supply unit 60. . Hereinafter, each part will be described.

The input / output section 40 is provided as an input section at the CCD camera 15 corresponding to the eyes of the robot apparatus 1, the microphone 16 corresponding to the ears, the head, the back, and the like to detect a user's contact. It includes a touch sensor 18, a paddle switch (not shown) disposed on each sole, or various other sensors corresponding to the five senses. Further, the output unit is equipped with a speaker 17 corresponding to a mouth, an LED indicator (eye lamp) 19 for forming a facial expression by a combination of blinking and lighting timing, and the like. These output units can express the user feedback from the robot apparatus 1 in a form other than the mechanical movement pattern by the legs such as voice and blinking of a lamp.

The drive unit 50 is a functional block that realizes a body motion of the robot apparatus 1 in accordance with a predetermined motion pattern instructed by the control unit 20, and is a control target under action control. The drive unit 50 is a functional module for realizing the degree of freedom at each joint of the robot apparatus 1, and is composed of a plurality of drive units provided for each axis such as roll, pitch, and yaw at each joint. Each drive unit adaptively controls the rotational position and rotational speed of the motor 51 based on the output of the encoder 51, the encoder 52 that detects the rotational position of the motor 51, and the rotational position of the motor 51. It is composed of a combination of drivers 53.

Depending on how the drive units are combined, the robot apparatus 1 can be configured as a legged mobile robot, for example, bipedal walking or quadrupedal walking.

The power supply unit 60 is a functional module that supplies power to each electric circuit and the like in the robot apparatus 1 as its name implies. The robot apparatus 1 according to the present embodiment is an autonomous drive type that uses a battery, and the power supply unit 60 includes a charging battery 61 and a charging / discharging control unit 62 that manages the charging / discharging state of the charging battery 61. .

The rechargeable battery 61 is constructed, for example, in the form of a "battery pack" in which a plurality of nickel-cadmium battery cells are packaged in a cartridge type.

The charging / discharging control unit 62 uses the battery 61
The remaining capacity of the battery 61 is grasped by measuring the terminal voltage, the amount of charging / discharging current, the ambient temperature of the battery 61, etc., and the charging start time, the charging end time, etc. are determined. The control unit 20 is notified of the start and end times of charging determined by the charge / discharge control unit 62, and serves as a trigger for the robot apparatus 1 to start and end the charging operation.

The control unit 20 corresponds to the "brain",
For example, it is mounted on the machine head or body of the robot apparatus 1.

FIG. 2 illustrates the configuration of the control unit 20 in more detail. As shown in the figure, the control unit 20 includes a CPU (Cent) as a main controller.
The ralProcessing Unit) 21 is bus-connected to a memory, other circuit components, and peripheral devices. The bus 27 is a common signal transmission line including a data bus, an address bus, a control bus and the like. A unique address (memory address or I / O address) is assigned to each device on the bus 27. The CPU 21 can communicate with a specific device on the bus 28 by designating an address.

A RAM (Random Access Memory) 22 is
It is a writable memory composed of a volatile memory such as a DRAM (Dynamic RAM), and is used for loading a program code executed by the CPU 21 and temporarily storing work data by the execution program.

A ROM (Read Only Memory) 23 is a read-only memory that permanently stores programs and data. The program code stored in the ROM 23 includes a self-diagnosis test program executed when the power of the robot apparatus 1 is turned on, and an operation control program that defines the operation of the robot apparatus 1.

The control program of the robot apparatus 1 includes a "sensor input / recognition processing program" which processes sensor inputs from the camera 15 and the microphone 16 and recognizes them as symbols, and storage operations such as short-term memory and long-term memory (described later). The "action control program" that controls the action of the robot apparatus 1 based on the sensor input and the predetermined action control model while controlling the drive of each joint motor and the voice output of the speaker 17 according to the action control model. Control program "and the like.

The non-volatile memory 24 is, for example, EEPRO.
It is composed of an electrically erasable and rewritable memory element such as M (Electrically Erasable and Programmable ROM), and is used for holding data to be sequentially updated in a non-volatile manner. The data to be sequentially updated includes an encryption key, other security information, a device control program to be installed after shipping, and the like.

The interface 25 is the control unit 2
It is a device that enables mutual data exchange by interconnecting with external devices. The interface 25 performs data input / output with the camera 15, the microphone 16, and the speaker 17, for example. Further, the interface 25 inputs and outputs data and commands with the drivers 53-1 ... In the drive unit 50.

Further, the interface 25 has an RS (Re
serial interface such as commended Standard) -232C, IEEE (Institute of Electrical an
d electronics Engineers) Parallel such as 1284
Interface, USB (Universal Serial Bus) interface, i-Link (IEEE1394) interface, SCSI (Small Computer System In)
terface), a memory card interface (card slot) that accepts a memory stick, etc., and a general-purpose interface for connecting peripheral devices of the computer, and programs and data with external devices that are locally connected. May be moved.

As another example of the interface 25, an infrared communication (IrDA) interface is provided,
Wireless communication may be performed with an external device.

Further, the control unit 20 includes a wireless communication interface 26, a network interface card (NIC) 27, etc., and is equipped with Bluetooth.
h proximity wireless data communication such as IEEE802.h.
Data communication can be performed with various external host computers via a wireless network such as 11b or a wide area network such as the Internet.

By such data communication between the robot apparatus 1 and the host computer, it is possible to calculate a complicated operation control of the robot apparatus 1 and to perform remote control using remote computer resources.

B. Action Control System of Robot Device FIG. 3 schematically shows a functional configuration of the action control system 100 of the robot device 1 according to the embodiment of the present invention. By associatively storing the change of the internal state from the external stimulus, the robot apparatus 1 can perform the action control according to the recognition result of the external stimulus and the change of the internal state.

Object-oriented programming can be adopted and implemented in the illustrated behavior control system 100. In this case, each piece of software is handled in module units called "objects" in which data and processing procedures for the data are integrated. In addition, each object uses a message communication method and an inter-object communication method using a shared memory to transfer data and Invoke.
e can be performed.

The behavior control system 100 uses the external environment (E
In order to recognize nvironments), a visual recognition function unit 101, an auditory recognition function unit 102, and a contact recognition function unit 103 are provided.

The visual recognition function unit (Video) 51 recognizes faces and colors based on a photographed image input through an image input device such as a CCD (Charge Coupled Device) camera 15. Image recognition processing such as is performed and feature extraction is performed, and a symbol corresponding to a visual recognition result of an object such as a face, a color, or a shape is output as an ID.

The auditory recognition function unit (Audio) 102 is
The voice data input via a voice input device such as the microphone 16 is voice-recognized to extract features and perform word set (text) recognition, and a symbol corresponding to the recognition result of the input voice is used as an ID. Output.

Contact recognition function unit (Tactile) 103
Is a symbol corresponding to the recognition result regarding the tactile sensation by recognizing a sensor signal from a contact sensor built in the head of the machine body or the like, recognizing an external stimulus such as “stroked” or “struck”. Is output as an ID.

Internal status management unit (ISM: Internal Statu
s Manager) 104 includes an instinct model and an emotional model, and includes the visual recognition function unit 101 and the auditory recognition function unit 1 described above.
02 and the internal state such as instinct and emotion of the robot apparatus 1 according to the external stimulus (ES: External Stimula) recognized by the contact recognition function unit 103.

The emotion model and the instinct model each have a recognition result and an action history as inputs, and manage emotion values and instinct values. The behavior model can refer to these emotional values and instinct values.

The robot apparatus 1 according to the present embodiment performs a behavior control according to the recognition result of the external stimulus and the change of the internal state, so that the robot apparatus 1 and the short-term storage unit 105 perform a short-term memory lost with the passage of time. A long-term storage unit 106 for holding information for a relatively long time is provided.

Short-term memory (ShortTermMemo
ry) 105 is a functional module that holds the targets and events recognized from the external environment by the above-described visual recognition function unit 101, auditory recognition function unit 102, and touch recognition function unit 103 for a short period of time. For example, the input image from the camera 15 is stored only for a short period of about 15 seconds.

Long-term memory (LongTermMemor
y) 106 is used to hold information obtained by learning such as the name of an object for a long period of time. Long-term memory 106
Can, for example, associatively store (described later) a change in internal state from an external stimulus in a certain action module.

The classification of memory mechanisms of short-term memory and long-term memory depends on neuropsychology, but details will be given later.

The behavior control of the robot apparatus 1 according to the present embodiment is performed by the reflexive behavior unit 109, the "reflexive behavior", the situation-dependent behavior hierarchy 108, the "situation-dependent behavior", and the deliberative behavior. It is roughly divided into "contemplation actions" realized by the hierarchy 107.

Reflexive Behavior Unit (Configuratio)
nDependentActionsAndReact
ions) 109 is the above-mentioned visual recognition function unit 101,
The auditory recognition function unit 102 and the contact recognition function unit 103 are functional modules that realize a reflexive body motion in response to an external stimulus recognized by the touch recognition function unit 103.

Basically, the reflex action is to directly receive the recognition result of the external information input by the sensor, classify the result,
It is an action that directly determines the output action. For example, behaviors such as chasing a human face or nodding are preferably implemented as reflexive behaviors.

Context-dependent behavior hierarchy (SituatedBe)
The layers 108 are the short-term storage unit 10.
5 and the stored contents of the long-term storage unit 106 and the internal state managed by the internal state management unit 104, the action corresponding to the situation where the robot apparatus 1 is currently placed is controlled.

The state-dependent action hierarchy 108 prepares a state machine for each action, classifies the recognition result of the external information inputted by the sensor, depending on the action and the situation before that, and classifies the action as a body. Expressed above. The situation-dependent action hierarchy 108 also realizes an action for keeping the internal state within a certain range (also referred to as “homeostasis action”). When the internal state exceeds the specified range, the internal state Activate the behavior so that it becomes easier for the behavior to return to within the range (actually, the behavior is selected in consideration of both the internal state and the external environment). Situation-dependent behavior has a slower reaction time than reflexive behavior.

Deliberation Behavior Hierarchy (Deliverative)
The Layer) 107 performs a relatively long-term action plan of the robot apparatus 1 based on the stored contents of the short-term storage unit 105 and the long-term storage unit 106.

The deliberative action is an action that is performed in accordance with a given situation or an instruction from a human being and makes a plan for realizing the reasoning. Such inference or planning may require processing time or calculation load rather than reaction time for the robot apparatus 1 to maintain the interaction, so that the above-mentioned reflexive behavior or situation-dependent behavior returns a reaction in real time, Reflective actions provide reasoning and planning.

C. Storage Mechanism of Robot Device As described above, the robot device 1 according to the present embodiment is
It has a short-term storage unit 105 and a long-term storage unit 106,
Such a memory mechanism relies on neuropsychology.

Further, this embodiment is a competitive neural network.
By applying the network as a memory model of the robot apparatus 1, all long-term memory is realized by the same engineering model, and additional learning in long-term memory can be realized. Also, by conducting statistical learning,
It is possible to realize long-term memory that is less affected by noise. In addition, by performing associative meaning storage, the meaning of an object can be associated with incomplete data. In addition, by performing associative episode memory,
The robot device can associate the stored place, time, and the like.

First, the types of storage will be described.

Short-term memory is literally a short-term memory and is lost over time. Short-term memory can be used for short-term retention of targets and events recognized from the external environment, such as sight, hearing, and touch.

The short-term memory further stores "sensory memory" in which sensory information (that is, the output from the sensor) is retained as a signal for about 1 second, and sensory memory is encoded and stored in a short-term with a limited capacity. It can be classified into "direct memory" and "working memory" in which situation changes and contexts are stored for several hours. Direct memory is said to be 7 ± 2 chunks, according to neuropsychological studies. Working memory is also called “intermediate memory” in contrast to short-term memory and long-term memory.

The long-term memory is used for holding information obtained by learning such as the name of an object for a long time. The same pattern can be processed statistically for robust storage.

Long-term memory is further classified into "declarative knowledge memory" and "procedural knowledge memory". Declarative knowledge memory
It consists of "episode memory," which is a memory of a scene (for example, a scene when taught), and "semantic memory", which is a memory of the meaning and common sense of words. Further, the procedural knowledge memory is a procedural memory such as how to use the declarative knowledge memory, and can be used to acquire an action for an input pattern.

As described above, episode memory is a type of long-term memory, and among them, it is also a type of declarative knowledge memory (also called assertive memory). For example, when thinking of riding a bicycle, remembering a scene (time, place, etc.) when riding a bicycle for the first time corresponds to episode memory.
After that, while the memory about the episode fades with the passage of time, the meaning is memorized. Also, the procedure for riding a bicycle is stored, which corresponds to procedural knowledge storage.

Generally, it takes time to store procedural knowledge. Whereas one can "say" by declarative knowledge memory, procedural knowledge memory is latent and manifests itself in the execution of actions.

The long-term memory according to this embodiment is realized by using an associative memory model using a competitive neural network.

The associative memory refers to a mechanism in which an input pattern consisting of a plurality of symbols is stored in advance as a memory pattern and a pattern similar to a certain one of them is recalled. FIG. 4 schematically shows the mechanism of the memory process in associative memory. Further, FIG. 5 schematically shows the mechanism of the recall process in associative memory. According to the associative memory, when a partially defective pattern is input, the closest stored pattern among the plurality of stored patterns can be output. This is because even when only an external stimulus consisting of incomplete data is given, it is possible to recall the meaning of an object by firing the corresponding neuron.

Associative memory is roughly classified into "self-associative associative memory" and "mutual-associative associative memory". The self-remembering type is a model in which a memorized pattern is directly drawn by a key pattern, and the mutual recollecting type is a model in which an input pattern and an output pattern are connected by a certain kind of associative relation.

In the present embodiment, the self-associative associative memory is adopted, which is easier to perform additional learning than the conventional memory model such as Hopfield or Associatron (described above). Memory is possible,
There are merits such as.

According to the additional learning, even if a new pattern is newly stored, the past memory is not overwritten and erased.

According to the statistical learning, the more the same thing is seen, the more it remains in memory, and if the same thing is repeatedly executed, it becomes hard to forget. In this case, in the storing process (see FIG. 4), even if a complete pattern is not input every time, it is repeatedly executed to converge on a large number of presented patterns.

C-1. Semantic memory This section describes an example in which the associative memory mechanism according to the present embodiment is applied to semantic memory.

The pattern remembered by the robot apparatus 1 (that is, the pattern input to the associative memory in FIG. 4) is
It is composed of a combination of an external stimulus to the robot apparatus 1 and an internal state.

Here, the external stimulus is the perceptual information obtained by recognizing the sensor input by the robot apparatus 1, and, for example, the color information and the shape information processed on the image input from the camera 15. , Face information, and more specifically,
It is composed of components such as color, shape, face, 3D general object, hand gesture, movement, voice, contact, smell and taste.
It

The internal state means, for example, instinct or emotion based on the body of the robot. Instinctive factors include, for example, fatigue, heat or body temperature.
e), pain, appetite or hunger, thirst, affection, curiosit
At least one of y), elimination, or sexuality. The emotional element is happiness (ha
ppiness, sadness, anger, surprise, disgust, fear, frustration, boredom, sleep (somnolen)
ce), sociability (gregariousness), patience,
Tense, relaxed, alert
ess, guilt, spite, honesty (loyalt)
at least one of y), submission or jealousy.

In the associative memory mechanism to which the competitive neural network according to this embodiment is applied, an input channel is assigned to each element forming these external stimuli and internal states. Further, each perceptual function module such as the visual recognition function unit 101 and the auditory recognition function unit 102 is
Rather than sending a raw signal that is a sensor output, the result of recognizing the sensor output is symbolized, and ID information (for example, color prototype ID, shape prototype ID, voice prototype ID, etc.) corresponding to the symbol is applied to the corresponding channel. It is designed to be sent to.

For example, each object segmented by the color segmentation module is added with a color prototype ID and input to the associative memory system. Further, the ID of the face recognized by the face recognition module is input to the associative memory system. Also,
The ID of the object recognized by the object recognition module is input to the associative system. In addition, the prototype ID of the word is input from the voice recognition module according to the user's utterance. At this time, the phoneme symbol string of the utterance (Phoneme Sequen
ce) is also input, so that the robot apparatus 1 can be made to speak by the process of memory / association. Regarding the instinct, an analog value can be handled (described later). For example, if the delta value of the instinct is stored in 80, it is possible to obtain an analog value of 80 by association.

Therefore, the associative memory system according to this embodiment can store external stimuli such as colors, shapes, voices, etc. or internal states as an input pattern consisting of a combination of symbolized IDs for each channel. it can.
That is, the associative memory system stores [color ID
Shape ID Face ID Voice ID ... Instinct ID (value) Emotion
ID].

The associative memory has a memory process and a recall process (see FIGS. 4 and 5). FIG. 6 shows the concept of the storage process of associative memory.

The memory pattern input to the associative memory system is composed of a plurality of channels assigned to each element of the external stimulus and the internal state (in the illustrated example, it is composed of 8 channels of input 1 to input 8). ). Then, ID information that symbolizes the recognition result of the corresponding external stimulus and the internal state is sent to each channel. In the illustrated example, it is assumed that the shading of each channel represents ID information. For example, when the k-th column in the memory pattern is assigned to the face channel, the color represents the face prototype ID.

In the example shown in FIG. 6, it is assumed that the associative storage system has already stored a total of n storage patterns 1 to n. Here, the difference in the color of the corresponding channel between the two storage patterns means that the symbols of external stimuli or internal states, that is, IDs, stored on the same channel are different between the storage patterns.

FIG. 7 shows the concept of the associative memory recall process. As described above, when a pattern similar to the input pattern stored in the storage process is input, a complete storage pattern is output so as to supplement the missing information.

In the example shown in FIG. 7, a pattern to which an ID is given only to the upper 3 channels of the memory pattern consisting of 8 channels is input as a key pattern. In such a case, the associative memory system finds a pattern (memory pattern 1 in the illustrated example) that is closest to these upper three channels among the memory patterns already stored and outputs it as a recalled pattern. be able to. That is, the missing channel 4
The nearest storage pattern is output so as to supplement the information of ~ 8.

Therefore, according to the associative memory system,
You can associate a voice ID, that is, a name from only the face ID,
You can recall "delicious" or "not delicious" from just the name of the food.

According to the long-term memory architecture based on the competitive neural network, it is possible to realize the semantic memory regarding the meaning of words and common sense with the same engineering model as other long-term memories.

C-2. Competitive neural network
According to the present embodiment, a competitive neural network is applied as a memory model of the robot apparatus 1 to realize all long-term memory by the same engineering model. Thereby, additional learning in long-term memory can be realized. In addition, by performing statistical learning, it is possible to realize long-term memory that is less affected by noise. Also,
By performing associative semantic memory, the meaning of a certain object can be associated with incomplete data. Further, by performing the associative episode storage, the robot device can associate the stored place and time.

FIG. 8 schematically shows a configuration example of an associative memory system to which a competitive neural network is applied. As shown in the figure, this competitive neural network has an input layer and a competitive layer.
It is a hierarchical neural network consisting of two layers (titive layer).

This competitive neural network is
It has two kinds of operation modes, a memory mode and an associative mode. In the memory mode, input patterns are stored competitively, and in the recall mode, a complete memory pattern is recalled from a partially missing input pattern.

The input layer is composed of a plurality of input neurons. A symbol corresponding to the recognition result of the external stimulus or the internal state, that is, ID information is input to each input neuron from a channel assigned to each element representing the external stimulus or the internal state. In the input layer, it is necessary to prepare the number of neurons corresponding to the number of color IDs + the number of shape IDs + the number of voice IDs + the kind of instinct.

The competitive layer is composed of a plurality of competitive neurons. Each competitive neuron is connected to each input neuron on the input layer side with a certain connection weight.
A competing neuron corresponds to one symbol that each neuron should store. In other words, the number of competitive neurons corresponds to the number of symbols that can be stored.

It is assumed that a certain input pattern is given to the input layer. At this time, the input pattern is made up of channels representing each element of the external stimulus and the internal state, and the input neuron to which the corresponding ID is sent from the channel fires.

The competing neuron weights the outputs from the respective input neurons with synapses and inputs them, and calculates the sum of these input values. Then, learning is performed by selecting a competitive neuron that maximizes the sum of input values in the competitive layer and strengthening the binding force between the competitive neuron that has won and the input neuron. In addition, by selecting a competitive neuron that has won in the competitive layer for a defective input pattern, the symbol corresponding to the input pattern can be recalled.

C-2-1. The weight of the storage mode input layer and the competition layer is assumed to be a value between 0 and 1. However, the initial connection weight is randomly determined.

The memory in the competitive neural network is performed by first selecting a competitive neuron that has won in the competitive layer for an input pattern to be stored and strengthening the coupling force between the competitive neuron and each input neuron.

Here, the input pattern vector [x ₁ ,
x ₂ , ..., X _n ] is a color prototype ID
Corresponding to 1, when ID1 is recognized, the neuron x ₁ is fired, and the shape and the voice are sequentially fired in that manner. The firing neuron has a value of 1, and the non-firing neuron has a value of -1.

When the coupling force between the i-th input neuron and the j-th competitive neuron is w _ij , the value of the competitive neuron y _j with respect to the input x _i is expressed by the following equation.

[0111]

[Equation 1]

Therefore, the neuron that wins out the competition can be obtained by the following equation.

[0113]

[Equation 2]

The memory is performed by strengthening the coupling force between the competitive neuron that has won in the competitive layer and each input neuron. Won neuron (winner n
euron) and the input neuron are updated by Koho
According to the update rule of nen, it is performed as follows.

[0115]

[Equation 3]

Here, normalization is performed with L2Norm.

[0117]

[Equation 4]

This binding force represents the so-called strength of memory,
It becomes memory.

Here, the learning rate α is a parameter representing the relationship between the number of times of presentation and the memory. The larger the learning rate α, the larger the weight is changed in one memory. For example, α =
If 0.5 is used, once memorized, it will not be forgotten, and if a similar pattern is presented next time, the memorized pattern can almost certainly be associated.

Further, the more it is presented and stored, the more it is stored.
The network connection value (weight) increases. This shows that the memory becomes stronger as the same pattern is input many times, statistical learning is possible, and long-term memory in which the influence of noise in the actual environment is small can be realized.

When a new pattern is input and an attempt is made to store it, a new neuron in the competitive layer is fired, so that the connection with the new neuron is strengthened and the connection with the neuron due to the previous memory is not weakened. Absent. In other words, in the associative memory by the competitive neural network according to this embodiment, additional learning is possible, and the problem of "forgetting" is released.

C-2-2. Recall Mode Now, the input pattern vector as shown below is shown in FIG.
Suppose that it is presented to the associative memory system shown in. The input pattern may not be complete but may be partially missing.

[0123]

[Equation 5]

At this time, the input vector may be the prototype ID, or the likelihood or probability for the prototype ID. The value of the output neuron y _j is calculated as follows for the input x _i .

[0125]

[Equation 6]

It can be said that the above equation represents the likelihood of the firing value of the competitive neuron according to the likelihood of each channel. What is important here is that for likelihood inputs from a plurality of channels, they can be connected to obtain the overall likelihood. In the present embodiment, only the ones associated with each other, that is, only the ones with the maximum likelihood are selected, and the neuron that wins out the competition can be obtained by the following equation.

[0127]

[Equation 7]

Since the obtained number of the competing neuron Y corresponds to the stored symbol number, the input pattern X can be recalled by the inverse matrix operation of W as shown in the following equation.

[0129]

[Equation 8]

C-3. Competitive neural network
Episode memory can be realized by using a long-term memory architecture based on a competitive neural network as shown in FIG.

Episode memory is a kind of long-term memory, and among them, declarative knowledge memory (also called assertive memory).
Is a kind of. For example, considering riding a bicycle,
Remembering the first time you ride a bicycle (time, place, etc.) is equivalent to episode memory.

FIG. 9 schematically shows the mechanism of the memory process in episode memory. As shown in the figure, the episode is input together with the sensory input pattern.
Then, an input pattern including the sensory input pattern and the episode pattern is stored as a memory pattern.

Here, the sensory input pattern is composed of a combination of an external stimulus to the robot apparatus 1 and an internal state. The external stimulus is perceptual information obtained by the robot device 1 recognizing a sensor input, and is a component such as color, shape, face, 3D general object, hand gesture, motion, voice, contact, smell, taste, or the like. Consists of. However, instead of directly using the signal detected by the sensor for the external stimulus, the result of recognizing the sensor output is symbolized, and ID information corresponding to the symbol (for example, color prototype ID, shape prototype ID, voice prototype ID, etc.) is generated. ).

An episode corresponds to a scene when it is stored, and is composed of constituent elements such as time, place, face, color and movement. These are treated as ID information and form an episode input pattern.

Episodes are composed of external stimuli and the like obtained at the time of memory. In order to acquire the episode input pattern based on the external stimulus, the robot apparatus 1 may include an episode detector. FIG. 10 schematically shows a configuration example of the episode detector.

First, the image at the time of storage is input to the episode detector. The episode detector requests other recognition function modules (not shown in FIG. 10) to recognize information other than time and place (self-position recognition). The face is information about who was with the face when the face was stored, and if there is a face other than the face to be stored in the image, its ID is returned to the episode detector as a face recognition result. Also, as for the color, an ID indicating which color is more visible in the image is returned to the episode detector. For example, when there are many reddish colors at dusk, a red ID is returned. Regarding the movement, the image of the previous frame is also necessary, but it can be obtained by taking the difference from the current image. For example, the movements are divided into stages such as “Yes”, “No”, and “Okay”, and an ID is given to each class. If there is a large difference from the previous frame, the ID is returned to the episode detector, such as motion.

The episode detector may assign an episode ID for specifying a scene to be memorized, such as ID = 0: morning ID = 1: daytime ID = 2: night. Alternatively, ID = 0: 0:00 to 1:00, ID = 1: 1:00 to 2:00 may be used.

FIG. 11 schematically shows the mechanism of the recall process in episode memory. As shown in the figure, when a sensory input pattern having a partial defect is input, the sensory memory pattern and the episode memory pattern that are the closest among a plurality of stored patterns can be output. That is, when a certain sensory pattern is input, an episode is also recalled, so that by listening to a certain word, it is possible to recall the situation or scene when it was remembered.

FIG. 12 schematically shows the configuration of a competitive neural network for episode memory. As shown in the figure, this competitive neural network has an input layer and a competitive layer.
This is a layered neural network consisting of two layers (ve layer).

The input layer is composed of a plurality of input neurons. A symbol corresponding to the recognition result of the external stimulus, that is, ID information is input to each input neuron from a channel assigned to each element representing the external stimulus. In addition to these, an episode detector (see FIG.
) Is provided for inputting an episode input pattern supplied from

In the input layer, in addition to the number of color IDs, the number of shape IDs, the number of voice IDs, the type of instinct, etc., it is necessary to prepare input neurons in which only the number of channels of the episode input pattern is added.

The competitive layer is composed of a plurality of competitive neurons. Each competitive neuron is connected to each input neuron on the input layer side with a certain connection weight.
A competing neuron corresponds to one symbol that each neuron should store. In other words, the number of competitive neurons corresponds to the number of symbols that can be stored.

In the memory in the competitive neural network shown in FIG. 12, first, a competitive neuron that has won in the competitive layer for the input pattern to be stored is selected, and the coupling force between the competitive neuron and each input neuron is selected. Do by strengthening.

Input patterns for competitive networks
・ The vector is [x₁, X₂, ..., x _n, X_{n + 1}，… ，
x_m] (Where n <m). Where x₁, X
₂, ..., x_nIs a sensory input neuron, and
x_{n + 1}, ..., x_mIs an episode input neuron
It Neuron x₁Corresponds to color prototype ID1,
If ID1 is recognized, neuron x₁Fire and order
Next, the shape and sound are also fired in that way. Similarly, the episode
Cord neuron x_{n + 1}Is the time prototype ID_{n + 1}Against
In response, ID_{n + 1}Is recognized, episode neuron
x_{n + 1}Ignite. The firing neuron has a value of 1.
Therefore, a neuron that does not fire has a value of -1.

When the coupling force between the i-th input neuron and the j-th competitive neuron is w _ij , the value of the competitive neuron y _j with respect to the input x _i is expressed by the following equation.

[0146]

[Equation 9]

Therefore, the neuron that wins out the competition can be obtained by the following equation.

[0148]

[Equation 10]

The memory is performed by strengthening the binding force between the competitive neuron that has won in the competitive layer (winner neuron) and each input neuron. Winning competitive neuron (winn
er neuron) and sensory input neurons and episodes
The update of the connection with the neuron is performed as follows according to the Kohonen update rule.

[0150]

[Equation 11]

Here, normalization is performed with L2Norm.

[0152]

[Equation 12]

This binding force represents the so-called strength of memory,
It becomes memory.

Here, the learning rate α is a parameter representing the relationship between the number of times of presentation and the memory. The larger the learning rate α, the larger the weight is changed in one memory. For example, α =
If 0.5 is used, once memorized, it will not be forgotten, and if a similar sensory input pattern or episode pattern is presented next time, the memorized episode pattern can almost certainly be associated.

Further, the more it is presented and stored, the more it is stored.
The network connection value (weight) increases. This indicates that the memory becomes stronger as the same episode pattern is input many times, statistical learning is possible, and long-term memory with less noise influence in a real environment can be realized. it can.

Further, if a new sensory input pattern and / or episode pattern is input and an attempt is made to memorize, a new neuron in the competitive layer is fired, so that the connection with the new neuron is strengthened, and the result of the previous memory is increased. It does not mean that the connection with the episode neuron is weakened. In other words, in the associative memory by the competitive neural network according to this embodiment, additional learning is possible, and the problem of "forgetting" is released.

[0157] Further, in the recall mode, sensory input pattern x ₁ as shown below, x _2, ..., x _n, and episodes pattern x _{n + 1,} ..., the input pattern vector of x _m Figure It is assumed that it is presented in the associative memory system shown in 12.

[0158]

[Equation 13]

At this time, the value of the output neuron y _j is calculated as follows for the input x _i .

[0160]

[Equation 14]

The above equation represents the likelihood of the firing value of the competitive neuron according to the likelihood of each channel. In the present embodiment, only the maximum likelihood is selected as an associative one, and the neuron that wins out the competition can be obtained by the following equation.

[0162]

[Equation 15]

Since the number of the competing neuron Y obtained corresponds to the number of the stored symbol, the input pattern X can be recalled by the inverse matrix operation of W as in the following equation.

[0164]

[Equation 16]

The input pattern X is the sensory input pattern [x
₁ , ..., _Xn ] and episode pattern [ _{xn + 1} , ...,
x _m ]. In other words, the closest sensory memory pattern and the closest episode memory pattern can be recalled based on the missing sensory input pattern.

A competitive neural network as shown in FIG.
According to the network long-term memory architecture,
It is possible to realize the episode memory about the scene when it is memorized by the same engineering model as other long-term memory.
By performing associative episode memory, that is, from an incomplete sensory input, an episode such as a place or time when a certain event is stored can be associated.

When a new pattern is input and an attempt is made to memorize, a new competing neuron is fired to strengthen the connection with the new neuron, but the connection of the neuron due to the previous memory is not weakened, so that additional learning is possible. Is.

Further, in the long-term memory architecture using the competitive neural network shown in FIG. 12, since the memory is strengthened as the same pattern is input many times, statistical learning is performed, so that the influence of noise is reduced. A small amount of long-term memory can be realized.

C-4. Competitive neural network
The procedural knowledge memory can be realized by using the long-term memory architecture by the competitive neural network as shown in FIG.

Long-term memory is further classified into "declarative knowledge memory" and "procedural knowledge memory". Declarative knowledge memory
It consists of "episode memory," which is a memory of a scene (for example, a scene when taught), and "semantic memory", which is a memory of the meaning and common sense of words. Further, the procedural knowledge memory is a procedural memory such as how to use the declarative knowledge memory, and can be used to acquire an action for an input pattern.

For example, when considering riding a bicycle, remembering the first riding scene (time, place, etc.) is equivalent to episode memory, and the procedure of riding a bicycle may be stored. Corresponds to procedural knowledge memory.

Memories of procedural knowledge are, as it were, the acquisition of verbs.

FIG. 13 schematically shows a storage process mechanism in procedural knowledge storage. As shown in the figure, the motion pattern is input together with the sensory input pattern. Then, an input pattern including the sensory input pattern and the motion pattern is stored as a storage pattern.

Here, the sensory input pattern is composed of a combination of an external stimulus to the robot apparatus 1 and an internal state. The external stimulus is perceptual information obtained by the robot device 1 recognizing a sensor input, and is a component such as color, shape, face, 3D general object, hand gesture, motion, voice, contact, smell, taste, or the like. Consists of. However, instead of directly using the signal detected by the sensor for the external stimulus, the result of recognizing the sensor output is symbolized, and ID information corresponding to the symbol (for example, color prototype ID, shape prototype ID, voice prototype ID, etc.) is generated. ).

The motion pattern is composed of the motion angle of each joint actuator and its angular velocity. Here, the motions of a plurality of joint actuators are combined and abstracted / symbolized, and treated as ID information, that is, motion ID. Therefore, when a certain motion on the body of the robot apparatus 1 occurs, a motion ID corresponding to the motion is sent to the corresponding channel of the input layer.

Further, FIG. 14 schematically shows the mechanism of the recall process in procedural knowledge memory. As shown in the figure, when a sensory input pattern having a partial defect is input, it is possible to output the closest sensory memory pattern and motion memory pattern among a plurality of stored patterns. That is, when a certain sensory pattern is input, the motion pattern is also recalled, and therefore, by listening to a certain word, it is possible to recall the motion when it is stored.

FIG. 15 schematically shows the structure of a competitive neural network for procedural knowledge storage. As shown in the figure, this competitive neural network has an input layer and a competitive layer.
This is a layered neural network consisting of two layers (ve layer).

The input layer is composed of a plurality of input neurons. A symbol corresponding to the recognition result of the external stimulus, that is, ID information is input to each input neuron from a channel assigned to each element representing the external stimulus. In addition to these, a motion input neuron for inputting a motion ID corresponding to a motion pattern demonstrated on the body of the robot apparatus 1 is provided.

In the input layer, in addition to the number of color IDs + the number of shape IDs + the number of voice IDs + the type of instinct ... is there.

The competitive layer is composed of a plurality of competitive neurons. Each competitive neuron is connected to each input neuron on the input layer side with a certain connection weight.
A competing neuron corresponds to one symbol that each neuron should store. In other words, the number of competitive neurons corresponds to the number of symbols that can be stored.

In the memory in the competitive neural network shown in FIG. 15, first, a competitive neuron that has survived in the competitive layer for an input pattern to be stored is selected, and the coupling force between the competitive neuron and each input neuron is selected. Do by strengthening.

Input patterns for competitive networks
・ The vector is [x₁, X₂, ..., x _n, X_{n + 1}，… ，
x_m] (Where n <m). Here here
And x₁, X ₂, ..., x_nIs a sensory input neuron,
Also, x_{n + 1}, ..., x_mIs a motion input neuron
It Neuron x₁Corresponds to color prototype ID1,
If ID1 is recognized, neuron x₁Fire and order
Next, the shape and sound are also fired in that way. Similarly,
Force neuron x_{n + 1}Is the operation prototype ID_{n + 1}Corresponding to
And ID_{n + 1}Is recognized, the motion input neuron x_{n + 1}To
Ignite. The firing neuron takes the value of 1 and fires
A neuron that does not has a value of -1.

If the coupling force between the i-th input neuron and the j-th competitive neuron is w _ij , the value of the competitive neuron y _j with respect to the input x _i is expressed by the following equation.

[0184]

[Equation 17]

Therefore, the neuron that wins out the competition can be obtained by the following equation.

[0186]

[Equation 18]

The memory is performed by strengthening the binding force between the competitive neuron that has won in the competitive layer (winner neuron) and each input neuron. Winning competitive neuron (winn
er neuron) and sensory input neurons and motion input neurons are updated as follows according to Kohonen's update rule.

[0188]

[Formula 19]

Here, normalization is performed with L2Norm.

[0190]

[Equation 20]

This binding force represents the strength of so-called memory,
It becomes memory.

Here, the learning rate α is a parameter indicating the relationship between the number of times of presentation and the memory. The larger the learning rate α, the larger the weight is changed in one memory. For example, α =
If 0.5 is used, once stored, it will not be forgotten, and if a similar sensory input pattern or motion input pattern is presented next time, the stored motion input pattern can almost certainly be associated.

[0193] Further, the more it is presented and stored, the more it is stored.
The network connection value (weight) increases. This indicates that the memory becomes stronger as the same motion input pattern is input many times, statistical learning is possible, and long-term memory with less influence of noise in a real environment can be realized. it can.

Further, if a new sensory input pattern and / or a motion input pattern is input and an attempt is made to memorize, a new neuron in the competitive layer is fired, so that the connection with the new neuron is strengthened, and the result of the previous memory is increased. The connection with the motion input neuron does not weaken. In other words, in the associative memory by the competitive neural network according to this embodiment, additional learning is possible, and the problem of "forgetting" is released.

[0195] Further, in the recall mode, sensory input pattern x ₁ as shown below, x _2, ..., x _n, and the operation input pattern x _{n +} 1, ..., the input pattern vector of x _m 15 Suppose that it is presented to the associative memory system shown in.

[0196]

[Equation 21]

At this time, the value of the output neuron y _j is calculated as follows for the input x _i .

[0198]

[Equation 22]

The above equation represents the likelihood of the firing value of the competitive neuron according to the likelihood of each channel. In the present embodiment, only the maximum likelihood is selected as an associative one, and the neuron that wins out the competition can be obtained by the following equation.

[0200]

[Equation 23]

Since the obtained number of the competitive neuron Y corresponds to the stored symbol number, the input pattern X can be recalled by the inverse matrix operation of W as in the following equation.

[0202]

[Equation 24]

The input pattern X is the sensory input pattern [x
_1, ..., x _n] and operation input pattern [x _{n +} 1, ..., consisting of x _m]. In other words, it is possible to recall the closest sensory memory pattern and the closest motion memory pattern based on the missing sensory input pattern. By inputting the evoked action pattern into the situation-dependent action hierarchy 108, the corresponding state machine is activated and the action evoked by the missing sensory input can be expressed.

The competitive neural network shown in FIG.
According to the network long-term memory architecture,
Procedural knowledge memory such as how to use declarative knowledge memory such as episodic memory and semantic memory can be realized with the same engineering model as other long-term memory. By performing the associative procedural knowledge storage, the stored procedure can be associated, that is, the operation can be executed from the incomplete sensory input. For example, if you look at a red and round object,
You can think of an apple and try to eat it.

When a new pattern is input and an attempt is made to memorize, a new competing neuron is fired and the connection with the new neuron is strengthened. However, since the connection of the neuron due to the previous memory is not weakened, additional learning is possible. Is.

Further, in the long-term memory architecture by the competitive neural network shown in FIG. 15, since the memory is strengthened as the same pattern is input many times, statistical learning is performed, so that the influence of noise is reduced. A small amount of long-term memory can be realized.

Memories of procedural knowledge are, so to speak, acquisition of verbs. When the competitive neural network is applied to procedural knowledge memory, in the recall mode, the closest motion memory pattern can be obtained based on the missing sensory input pattern. Therefore, it can be configured to induce the action of the recollected aircraft by the procedural knowledge association.

In this case, as shown in FIG. 16, the motion pattern evoked by the neural network may be directly input to the behavior control section (state machine) to drive and control the machine body.

As an example, suppose that a "red" and "round" thing is named "apple", and that the user eats it. However, ID 1 is assigned to the action of “eat”.

First, when memorizing, "red" ID, "round" ID, "apple" ID, "eat" ID together,
It is stored in the long-term storage system as shown in FIG.

Next, in the recall mode, it is assumed that a red and round sensation input pattern is input. In response, the competitive neural network uses the name "apple" and "eat" by associative memory.
Recalling the above motion, it is acquired as the closest sensory memory pattern and the closest motion memory pattern.

By transmitting the ID "eat" to the state machine that manages the action, the robot apparatus can take the action of eating only by seeing the red and round object.

FIG. 17 shows a modification of the competitive neural network shown in FIG. 15 for procedural knowledge storage.

The neural network shown in FIG.
The input layer is composed of a plurality of input neurons and receives a sensory input pattern. The input layer is composed of a single layer, and the competitive layer is composed of two layers of competitive layers for associative memory.

One of the competitive layers is the noun layer, which mainly stores the noun part of the input pattern in an associative manner. Also,
The other competing layer is the verb layer, which mainly associates and memorizes the verb portion of the input pattern.

The output of the noun layer shows the ID of the noun. The output of the verb layer shows the ID of the verb,
The action control section can output the operation to the drive section.

When a sensory input pattern is input to the input layer, first, the first competitive layer competes to acquire a noun by the winning competitive neuron. Furthermore, with this winning noun neuron, they compete in the second layer of the competitive layer,
The verb neuron, which has the strongest connection with the competitive neuron that won in the first layer, fires.

A verb ID is given to each neuron in the second layer of the competitive layer. Therefore, the action control unit (state machine) may output the action according to the output verb ID.

As shown in FIG. 17, the efficiency of the calculation can be improved by dividing the calculation into the noun layer and the verb layer. Also, when the number of verbs increases, it is necessary to tune between the noun neuron and the verb neuron when it is composed of a single competitive layer, but when the noun layer and the verb layer are independent. Is free from such complications.

[Supplement] The present invention has been described in detail with reference to the specific embodiments. However, it is obvious that those skilled in the art can modify or substitute the embodiments without departing from the scope of the present invention.

In the above-mentioned embodiments, the case where the associative memory mechanism according to the present invention is applied to the robot apparatus has been mainly described. The gist of the present invention is not necessarily a “robot”
Is not limited to the products referred to as. That is, as long as it is a mechanical device that performs a motion similar to a human motion by using an electric or magnetic action, even if it is a product belonging to another industrial field such as a toy, the present invention is similarly applied. Can be applied.

In short, the present invention has been disclosed in the form of exemplification, and the contents described in this specification should not be construed in a limited manner. To determine the gist of the present invention,
The claims section mentioned at the beginning should be taken into consideration.

[0223]

As described above in detail, according to the present invention,
An excellent robot apparatus capable of appropriately associatively storing a change in an internal state from an external stimulus, and a behavior control method thereof,
An associative memory device and an associative memory method can be provided.

Further, according to the present invention, the information obtained by learning such as the name of an object can be retained for a long time by associatively storing the change of the internal state from the external stimulus by using the neural network, which is excellent. A robot device, a behavior control method thereof, an associative memory device, and an associative memory method can be provided.

Further, according to the present invention, an excellent robot apparatus capable of performing additional learning and capable of strong learning against noise in a real environment, an action control method therefor, an associative memory apparatus and an associative memory method are provided. Can be provided.

According to the present invention, engineering models of all long-term memories such as semantic memory, episodic memory, and procedural knowledge memory can be constructed within the same framework of competitive neural networks.

The associative memory system according to the present invention can realize additional learning in long-term memory by applying the competitive neural network to the memory model. In addition, by performing statistical learning, it is possible to realize long-term memory that is less affected by noise. Also,
By performing associative semantic memory, the meaning of a certain object can be associated with incomplete data. Further, by performing the associative episode storage, the robot device can associate the stored place and time.

Further, according to the associative memory system of the present invention, by applying the competitive neural network to the memory model, procedural knowledge memory is performed and the verb at the ID level is acquired from the missing sensory input pattern. Is possible. For example, you can associate the person's name just by looking at the face. Also, when you see a red and round object, you can control the behavior based on associative memory, such as "Let's eat because it is an apple."

Learning in a mobile robot device such as a legged robot is affected by noise because it operates in a real environment. According to the present invention, since the memory is stored statistically, effective and robust memory can be realized.

[Brief description of drawings]

FIG. 1 is a diagram schematically showing a functional configuration of a robot device 1 used for implementing the present invention.

FIG. 2 is a diagram showing the configuration of a control unit 20 in more detail.

FIG. 3 is a diagram schematically showing a functional configuration of a behavior control system 100 of the robot device 1 according to the embodiment of the present invention.

FIG. 4 is a diagram schematically showing a mechanism of a memory process in associative memory.

FIG. 5 is a diagram schematically showing a mechanism of a recall process in associative memory.

FIG. 6 is a diagram conceptually showing a mechanism of a memory process in associative memory.

FIG. 7 is a diagram conceptually showing a mechanism of a recall process in associative memory.

FIG. 8 is a diagram schematically showing a configuration of a competitive neural network.

FIG. 9 is a diagram schematically showing a mechanism of a memory process in episode memory.

FIG. 10 is a diagram schematically showing a configuration example of an episode detector.

FIG. 11 is a diagram schematically showing a mechanism of a recall process in episode memory.

FIG. 12: Competitive neural network for episode memory
It is the figure which showed the structure of the network typically.

FIG. 13 is a diagram schematically showing a storage process mechanism in procedural knowledge storage.

FIG. 14 is a diagram schematically showing a mechanism of a recall process in procedural knowledge memory.

FIG. 15: Competitive neural network for procedural knowledge storage
It is the figure which showed the structure of the network typically.

FIG. 16 is a diagram schematically showing a mechanism of a recall process in which an action is induced by procedural knowledge association.

17 is a diagram showing a modification of the competitive neural network for performing procedural knowledge storage shown in FIG.

[Explanation of symbols]

1 ... Mobile robot 15 ... CCD camera 16 ... Microphone 17 ... speaker 18 ... Touch sensor 19 ... LED indicator 20 ... Control unit 21 ... CPU 22 ... RAM 23 ... ROM 24 ... Non-volatile memory 25 ... Interface 26 ... Wireless communication interface 27 ... Network Interface Card 28 ... bus 29 ... Keyboard 40 ... Input / output section 50 ... Drive unit 51 ... Motor 52 ... Encoder 53 ... Driver 100 ... Behavior control system 101 ... Visual recognition function unit 102 ... Auditory recognition function unit 103 ... Contact recognition function unit 105 ... Short-term memory 106 ... Long-term storage 107 ... Contemplation stage 108 ... Situation-dependent behavior hierarchy 109 ... Reflex action section

   ─────────────────────────────────────────────────── ─── Continued front page    F-term (reference) 3C007 AS36 CS08 JS03 KS31 KS39                       KT01 LW12 LW15 WA13 WA14                       WB14 WB15 WB16 WB19 WB25                 5H004 GA15 GA17 GA26 GB16 HB03                       HB15 JB06 JB09 KD62 KD63                       MA05 MA06 MA23 MA36 MA39                       MA53 MA54 MA55 MA56

Claims (28)

[Claims] 1. A drive unit for performing joint drive or interactive drive with a user, an associative storage unit for storing a predetermined input pattern and for recalling a complete memory pattern from a missing input pattern, and an association memory unit for recalling. An action control unit that controls the operation of the aircraft by the drive unit based on the stored memory pattern,
A robot apparatus comprising: 2. The associative memory unit comprises an input layer composed of a plurality of input neurons that fires according to an input pattern, and a competition composed of a plurality of competing neurons that weight-input the outputs from the respective input neurons according to the coupling force. The robot apparatus according to claim 1, wherein associative memory is performed using a competitive neural network having layers. 3. The associative memory unit selects a competitive neuron having a maximum sum of weighted input values from each input neuron with respect to a certain input pattern, and determines a coupling force between the competitive neuron and each input neuron. The robot apparatus according to claim 2, wherein the memory of the input pattern is enhanced by strengthening the input pattern. 4. The associative memory unit selects a competitive neuron having the maximum sum of weighted input values from each input neuron for an input pattern in which a part is missing, and has a strong binding force with the competitive neuron. The robot apparatus according to claim 2, wherein a memory pattern including an input neuron is recalled. 5. A sensation input unit for inputting an external stimulus is further provided, wherein the associative storage unit stores the sensation input pattern from the sensation input unit and complete sensation memory pattern from the deficiency sensory input pattern. The robot apparatus according to claim 1, wherein: 6. A sensory input section for inputting an external stimulus, and an internal state management section for managing an internal state consisting of instinct and emotion, wherein the associative memory section receives sensory input and internal state from the sensory input section. The robot apparatus according to claim 1, wherein the robot apparatus stores an input pattern consisting of the following and recalls a complete sensory memory pattern and / or an internal state pattern from a sensory input pattern that is partially missing. 7. A sense input unit for inputting an external stimulus, and an episode detection unit for detecting an episode consisting of time, place, face, color, movement, etc. stored based on the external environment, and the associative storage unit. Storing a sensory input from the sensory input unit and an input pattern composed of detected episodes, and recalling a complete sensory memory pattern and / or episode pattern from the sensory input pattern partially missing. The robot apparatus according to claim 1. 8. A sensation input unit for inputting an external stimulus is further provided, wherein the associative storage unit stores and partially stores an input pattern consisting of a sensation input from the sensation input unit and an operation executed in the drive unit. Recalls a complete sensory memory pattern and / or an operation pattern of the drive unit from the sensory input pattern in which the action control unit has a loss, and the action control unit controls the aircraft operation according to the action pattern of the drive unit that is recalled. The robot apparatus according to claim 1, wherein the robot apparatus is a robot apparatus. 9. The competitive layer comprises a noun layer composed of a plurality of noun neurons storing nouns corresponding to input patterns,
A noun neuron that has a verb layer composed of a plurality of verb neurons that stores a verb corresponding to each noun, and first acquires a noun by a noun neuron that has competed in a noun layer for a certain input pattern 3. The robot apparatus according to claim 2, wherein the verb neuron having the strongest connection with the noun neuron fires by competing in the verb layer. 10. A method for controlling a behavior of a robot apparatus for driving a joint or driving a dialogue with a user, comprising: associative memory step of storing a predetermined input pattern and recalling a complete memory pattern from a missing input pattern. A behavior control step of controlling a body motion of the robot apparatus based on a memory pattern recollected by the associative memory step, the behavior control method of the robot apparatus. 11. The associative memory step comprises an input layer composed of a plurality of input neurons that fires according to an input pattern, and a competition composed of a plurality of competing neurons that weight-input the outputs from each input neuron according to the coupling force. 11. The behavior control method for a robot apparatus according to claim 10, wherein associative memory is performed using a competitive neural network having layers. 12. In the associative memory step, a competitive neuron having the maximum sum of weighted input values from each input neuron is selected for a certain input pattern, and the coupling force between the competitive neuron and each input neuron is selected. 12. The behavior control method for a robot apparatus according to claim 11, wherein the memory of the input pattern is enhanced by strengthening the input pattern. 13. In the associative memory step, a competitive neuron having a maximum sum of weighted input values from each input neuron is selected for an input pattern in which a part is missing, and a strong binding force with the competitive neuron is selected. The behavior control method for a robot apparatus according to claim 11, wherein a memory pattern including an input neuron is recalled. 14. A sensation input step of inputting an external stimulus is further provided, wherein in the associative memory step, the sensation input pattern in the sensation input step is stored and a complete sensation memory pattern is extracted from a deficient sensation input pattern. The action control method for a robot apparatus according to claim 10, wherein the action control method is recalled. 15. The method further comprises: a sensory input step of inputting an external stimulus; and an internal state management step of managing an internal state consisting of instinct and emotion, wherein the associative memory step is performed from the sensory input and the internal state in the sensory input step. 11. The behavior control method for a robot apparatus according to claim 10, further comprising storing the following input pattern and recalling a complete sensory memory pattern and / or an internal state pattern from the sensory input pattern partially missing. 16. A sense input step of inputting an external stimulus, and an episode detection step of detecting an episode consisting of time, place, face, color, movement, etc. stored based on the external environment, further comprising: And storing an input pattern consisting of the sensory input and the detected episode in the sensory input step and recalling a complete sensory memory pattern and / or episode pattern from the sensory input pattern partially missing. The behavior control method of the robot apparatus according to claim 10. 17. A sensation input step of inputting an external stimulus, wherein the associative storage step stores the sensation input in the sensation input step and an input pattern consisting of a motion executed on a body of the robot apparatus. It is characterized in that a complete sensory memory pattern and / or a motion pattern of the machine body is recalled from a partially missing sensory input pattern, and in the action control step, the machine body motion is controlled according to the recalled motion pattern. The behavior control method of the robot apparatus according to claim 10. 18. The competitive layer comprises a noun layer composed of a plurality of noun neurons storing nouns corresponding to an input pattern, and a verb layer composed of a plurality of verb neurons storing verbs corresponding to each noun. First, a noun neuron that competes in a noun layer for a certain input pattern first wins a noun, and the acquired noun neuron competes in a verb layer to fire a verb neuron most strongly coupled to the noun neuron. The action control method for a robot apparatus according to claim 11, wherein: 19. An input layer composed of a plurality of input neurons firing according to an input pattern, and a competitive layer composed of a plurality of competing neurons for weighting and inputting the outputs from the respective input neurons according to the coupling force. Using a competitive neural network, selecting a competitive neuron that maximizes the sum of weighted input values from each input neuron for a certain input pattern and strengthening the binding force between the competitive neuron and each input neuron A memory pattern composed of an input neuron that has a strong coupling force with the competitive neuron by selecting the competitive neuron that maximizes the sum of the weighted input values from each input neuron with respect to the partially missing input pattern. An associative memory device that recalls at least some of the input neurons. An associative memory device that is assigned with intense sensory input, stores the sensory input pattern, and recalls a complete sensory memory pattern from the sensory input pattern that is partially missing. 20. Another part of the input neuron is assigned to an internal state composed of instinct and emotion, and stores an input pattern composed of sensory input and internal state, and stores a part of the sensory input pattern lacking. The associative memory device according to claim 19, wherein the associative memory device recalls a complete sensory memory pattern and / or an internal state pattern. 21. Another part of the input neuron is assigned to an episode consisting of stored time, place, face, color, movement, etc., and stores an input pattern consisting of sensory input and episode, and a part thereof. 20. The associative memory device according to claim 19, wherein a complete sensory memory pattern and / or episode pattern is recalled from the sensory input pattern in which is deleted. 22. Another part of the input neuron is assigned to an executed motion pattern, stores an input pattern consisting of a sensory input and an executed motion, and stores a part of the sensory input pattern from a missing sensory input pattern. The associative memory device according to claim 19, wherein the associative memory device recalls a complete sensory memory pattern and / or a motion pattern. 23. The competitive layer comprises a noun layer composed of a plurality of noun neurons storing nouns corresponding to input patterns, and a verb layer composed of a plurality of verb neurons storing verbs corresponding to each noun. First, a noun neuron that competes in a noun layer for a certain input pattern first wins a noun, and the acquired noun neuron competes in a verb layer to fire the verb neuron most strongly coupled to the noun neuron. 20. The associative memory device according to claim 19, wherein. 24. An input layer composed of a plurality of input neurons that fires according to an input pattern, and a competition layer composed of a plurality of competitive neurons that inputs the outputs from the respective input neurons with weights according to the coupling force. An associative memory method using a competitive neural network, wherein a competitive neuron having the maximum sum of weighted input values from each input neuron for a certain input pattern is selected and the competitive neuron and each input neuron are selected. The memory step for increasing the binding force of the input neuron and the input neuron with strong binding force to the competitive neuron by selecting the competitive neuron with the maximum sum of the weighted input values from each input neuron for the partially missing input pattern. And a recall step of recalling a memory pattern consisting of The part is assigned a sensory input by an external stimulus, and stores the sensory input pattern in the memory step, and recalls a complete sensory memory pattern from the partially missing sensory input pattern in the recall step. Associative memory method. 25. Another part of the input neuron is assigned to an internal state composed of instinct and emotion, and the memory step stores an input pattern composed of sensory input and internal state, 25. The associative memory method according to claim 24, wherein a complete sensory memory pattern and / or an internal state pattern is recalled from a sensory input pattern in which a part is missing. 26. Another part of the input neuron is assigned to an episode consisting of stored time, place, face, color, movement, etc., and the storing step stores an input pattern consisting of sensory input and episode. At the same time, the recall step recalls a complete sensory memory pattern and / or episode pattern from the sensory input pattern partially missing.
25. The associative storage method according to claim 24, wherein: 27. Another part of the input neuron is assigned to an executed action pattern, the storing step stores an input pattern consisting of a sensory input and an executed action, and the recalling step stores one input pattern. 25. The associative memory method according to claim 24, wherein a complete sensory memory pattern and / or motion pattern is recalled from the sensory input pattern in which a part is missing. 28. The competitive layer includes a noun layer composed of a plurality of noun neurons storing a noun corresponding to an input pattern, and a verb layer composed of a plurality of verb neurons storing a verb corresponding to each noun. In the recall step, a noun neuron that first competes in a noun layer for a certain input pattern and wins is used to acquire a noun, and the acquired noun neuron competes in a verb layer to have the strongest connection with the noun neuron. The verb neuron fires.
4. The associative memory method described in 4.