Community Research and Development Information Service - CORDIS


This paper presents an architecture that solves three issues that any reinforcement-based reactive system should face to be of real application. The first issue is to determine where to search the suitable action to handle each situation. The second one is to speed up the learning phase, and the third issue is to build a system with powerful generalisation abilities. The proposed architecture is based on: (i) the integration of reaction and planning in order to explore efficiently the action space in real time; (ii) the extension of the reinforcement connectionist framework with one-shot learning for codifying immediately promising rules; (iii) the building of the reactive component as a modular network in order to avoid negative interferences when learning different sets of rules; (iv) the use of constructive techniques for determining the suitable size for each module. This architecture is the core of a path-finder which deals with continuous-valued inputs and outputs, learns extremely quickly, and exhibits good noise-tolerance and generalisation capabilities.

Additional information

Authors: MILLAN J DEL R, JRC Ispra (IT)
Bibliographic Reference: Paper presented: 10th European Conference on Artificial Intelligence, Wien (AT), August 3-7, 1992
Availability: Available from (1) as Paper EN 36885 ORA
Record Number: 199210827 / Last updated on: 1994-12-02
Original language: en
Available languages: en