MAGNets: Micro-Architectured Group Neural Networks

Reinforcement Learning (RL) algorithms have successfully achieved human-like performances in complex environments like games, autonomous vehicles, and industrial robots. However, the Deep Neural Networks (DNNs) used to approximate large Deep Reinforcement Learning (DRL) policies are resource-hungry and opaque. This limits the applicability of DRL in safety-critical applications running on resource-constrained platforms. On the other hand, on inspecting the design of most multi-output safety critical embedded control applications, it may be observed that such systems often derive each output based on some artifacts, which are, in turn, derived from input variables. Given such dependencies of internal system states on inputs, one may argue that each of these derived artifacts can be approximated by a smaller network in a multi-network DRL setting. In this work, we propose Micro Architecture Group Neural Networks (MAGNets) that can distill the learning of a large DRL network into multiple small neural networks. Using several OpenAI Gym environments, we show that existing verification tools can be used to verify the output of MAGNets while preserving the performance of a large neural policy. We also report our gains in network compactness, which directly impacts the execution latency and applicability in edge devices.