Reward effect in reinforcement learning systems

Abstract EN

Learning Classifier Systems (LCS), are a machine learning technique which combines reinforcement learning, evolutionary computing and other heuristics to produce adaptive systems.

The system HRC (Human –Rat- Cheese) focuses in creating artificial creature (Rat) using computer simulation, and learning it how to choose between two different basic behaviors, (approach / escape) combining them to perform complex behavior, which represents the final response in changing environment.

The HRC is built of two-classifier subsystems working together, each classifier system learns a simple behavior, and the system as a whole has as its learning goal the control of activities.

Flat architecture was used.

The flat organization allows distinguishing between two different learning activities : the learning of basic behavior and the learning of switch behavior.

One classifier system learns basic behavior, (approach / escape), i.e., it is used to learn the simulated robot single step movement in every direction in the environment.

Whereas the other classifier system learns to control the activities of basic classifier systems, i.e., it is used to learn to choose between basic behaviors using suppression as a composition mechanism to chose between two basic behaviors which represent complex behavior.

Simple experiments were executed for HRC : comparing and contrasting the effect of the reinforcement learning using reward & punishment with learning using reward only.

Experiment results show that the run using reinforcement learning with reward only is unable to perform as well as the run with reinforcement learning with reward and punishment.