Collective learning systems: a model for automatic control

The author investigates how CLS (collective learning system) theory can be used to model and implement adaptive control systems. The pole-balancing problem is posed as the experimental system paradigm. In the CLS approach, collective learning automata acquire knowledge through a learning process consisting of trial-and-error interactions with the environment. It was demonstrated that, using very simple evaluation, selection, and update functions, the CLA (collective learning automation) was able to learn satisfactory control of the cart-pole system. The success rate of 85% at the end of the learning period was mainly the result of a relatively low incidence of the states corresponding to starting angles close to 60 and 120 degrees. The performance can be improved by presenting the CLA with a comprehensive range of operating conditions