Autonomous agent based on reinforcement learning and adaptive shadowed network

Abstract

The planning of intelligent robot behavior plays an important role in the development of flexible automated systems. The robot’s intelligence comprises its capability to act in unpredictable and chaotic situations, which requires not just a change but the creation of the robot’s working knowledge. Planning of intelligent robot behavior addresses three main issues: finding task solutions in unknown situations, learning from experience and recognizing the similarity of problem paradigms. This article outlines a planning system which integrates the reinforcement learning method and a neural network approach with the aim to ensure autonomous robot behavior in unpredictable working conditions.

The assumption is that the robot is a tabula rasa and has no knowledge of the work space structure. Initially, it has just basic strategic knowledge of searching for solutions, based on random attempts, and a built-in learning system. The reinforcement learning method is used here to evaluate robot behavior and to induce new, or improve the existing, knowledge. The acquired action (task) plan is stored as experience which can be used in solving similar future problems. To provide the recognition of problem similarities, the Adaptive Fuzzy Shadowed neural network is designed. This novel network concept with a fuzzy learning rule and shadowed hidden layer architecture enables the recognition of slightly translated or rotated patterns and does not forget already learned structures.

The intelligent planning system is simulated using object-oriented techniques and verified on planned and random examples, proving the main advantages of the proposed approach: autonomous learning, which is invariant with regard to the order of training samples, and single iteration learning progress.