A NEURAL ARCHITECTURE FOR ONLINE PATH LEARNING IN MAZE NAVIGATION

摘要

This paper describes a neural network architecture and the online learning policies that permits to an autonomous robot navigates though a maze in order to memorize a path that explores the entire environment, while avoiding obstacles. The state space representation is constructed by unsupervised and competitive learning as well as the mapping state-action is constructed by means of reinforcement learning, during the maze exploration. The result of learning creates a memory of states-actions that emerges an intelligent behavior, such as the path learning. The robot uses only its own infrared distance-sensors to perform obstacle detection, used as pattern recognition cues, while moving in a maze environment. In order to demonstrate the effectiveness and real-time ability of the proposed neural controller, we report a number of simulation results of navigation in unknown maze environments.