Biologically-inspired dynamic reconstruction of perceptual states using concept formation technique

The sources of intelligence for primitively facilitated robotic agents exist in their reperceptualizing capability of what they physically sense from the environment. To implement this capability we adopt an inductive learning method of conceptual formation for adaptively organizing a state space and develop a new algorithm for a robot to construct its task-relevant state space efficiently through matching encountering states with similar situations in the past and through generalizing them. We propose a methodology to dynamically increase the resolution of state spaces both adaptively and selectively by applying a concept formation technique from machine learning in a recursive fashion to a record of a sensorimotor history of a learning agent. By connecting this with the conventional reinforcement learning we showed our algorithm can perform tasks without suffering from hidden state problems in an artificial maze environment and also present its robustness even for a robot whose perceptual resources are quite restricted and/or bounded.