Naslov
Improving Search Effi ciency in the Action Space of an Instance-Based Reinforcement Learning Technique for Multi-robot Systems

Autori
Yasuda, Toshiyuki ; Ohkura, Kazuhiro

Vrsta, podvrsta i kategorija rada
Poglavlja u knjigama, znanstveni

Knjiga
Advances in Artificial Life

Urednik/ci
F. Almeida e Costa et al.

Izdavač
Springer

Grad
Berlin : Heidelberg

Godina
2007

Raspon stranica
325-334

ISBN
978-3-540-74912-7

Ključne riječi
Multi-robot System, Reinforcement Learning, Autonomous

Sažetak
We have developed a new reinforcement learning technique called Bayesian-discrimination-function-based reinforcement learning (BRL). BRL is unique, in that it not only learns in the predefi ned state and action spaces, but also simultaneously changes their segmentation. BRL has proven to be more eff ective than other standard RL algorithms in dealing with multi-robot system (MRS) problems, where the learning environment is naturally dynamic. This paper introduces an extended form of BRL that improves its learning effi ciency. Instead of generating a random action when a robot encounters an unknown situation, the extended BRL generates an action calculated by a linear interpolation among the rules with high similarity to the current sensory input. In both physical experiments and computer simulations, the extended BRL showed higher search effi ciency than the standard BRL.

Izvorni jezik
Engleski

Znanstvena područja
Strojarstvo

POVEZANOST RADA