Reinforcement Learning Based on a Bayesian Confidence Propagating Neural Network

• We present a system capable of reinforcement learning (RL) based on the Bayesian confidence propagating neural network (BCPNN). The system is called BCPNNRL and its architecture is somewhat motivated by parallels to biology. We analyze the systems properties and we benchmark it against a simple Monte Carlo (MC) based RL algorithm, pursuit RL methods, and the Associative Reward Penalty (AR-P) algorithm. The system is used to solve the n-armed bandit problem, pattern association, and path finding in a maze.

Ämnesord

NATURVETENSKAP  -- Data- och informationsvetenskap -- Datavetenskap (hsv//swe)

NATURAL SCIENCES  -- Computer and Information Sciences -- Computer Sciences (hsv//eng)

Publikations- och innehållstyp

ref (ämneskategori)

kon (ämneskategori)