Infusing common-sensical prior knowledge into topological representations of learning robots

Kana Ogawa, Pitoyo Hartono

Abstract: Good initializations are always beneficial for efficiently training neural networks. In this study, we propose a method for initializing a layered neural network with common-sensical human knowledge in an intuitive manner. To make this idea possible, a hierarchical neural network containing a two-dimensional topological map is adopted as an adaptive controller for physical robots. The low dimensionality of the map allows humans to hand-design the topological structure to infuse human common-sensical knowledge. The prior knowledge initialization subsequently allows the robot to execute efficient reinforcement learning in a real environment. The primary novelty of this study is the idea for transferring human common-sensical knowledge that may be subjective and not necessarily logically or mathematically expressible. While transfer learning between two different neural networks is common, transfer learning from human to neural networks is not sufficiently studied, and hence, this study offers a new idea for collaborative learning between human and neural networks. The idea is tested in the context of the real-time reinforcement learning of real robots.