Portal de Periódicos da CAPES

A reinforcement connectionist approach to robot path finding in non-maze-like environments

1992; Springer Science+Business Media; Volume: 8; Issue: 3-4 Linguagem: Inglês

10.1007/bf00992702

1573-0565

José del R. Millán, Carme Torras,

Reinforcement Learning in Robotics

This paper presents a reinforcement connectionist system which finds and learns the suitable situation-action rules so as to generate feasible paths for a point robot in a 2D environment with circular obstacles. The basic reinforcement algorithm is extended with a strategy for discovering stable solution paths. Equipped with this strategy and a powerful codification scheme, the path-finder (i) learns quickly, (ii) deals with continuous-valued inputs and outputs, (iii) exhibits good noise-tolerance and generalization capabilities, (iv) copes with dynamic environments, and (v) solves an instance of the path finding problem with strong performance demands.