We presented a hybrid deep learning architecture by combining biological brain-inspired Neural Circuit Policies with convolutional neural architecture. This allows the agent to learn key coherent features from various environments, express generalizability and interpret dynamics. We found that a combination of 12 inter-neurons and 8 command neurons can effectively perform high-stakes decision-making tasks such as autonomous vehicles. Our architecture not only has 305 times fewer trainable parameters than competing approaches but shows superior performance, robustness, and noise resiliency in famous autonomous vehicle simulations i.e. OpenAI-CarRacing and Udacity Simulator.
Make this part of your paper trail
Save this paper to a shelf, write a review, and keep your own notes.