Action evolution

Action evolution (AE) is a technique used in the fields of (UTC)

Reinforcement learning can be divided into policy reinforcement and action reinforcement . Policy reinfocement is the more common of the two, involving methods like Dinamic programming, Monte Carlo method & Temporal difference. In these methods optimal control policies are discovered and reinforced.

Action reinforcement methods usually applies genetic algorithms (GA) to doscover optimal policies.

AE implies the use of evolutionary methods for action reinforcement. This is achieved by introducing an intermediate stage of finding feasible actions for a given state. A GA explores the critic to find such actions. Finding this action is usually a problem of much less dimension than finding the full set of policy parameters. It further offers the parallel global search benefits of GAs with immediate fitness evaluation. These actions can be assessed instantaneously and do not require any trial period as is the case with policy reinforcement. The results are recorded as state-action data pairs for supervised learning of the actor.

= Reference =

*Action Evolution for Intelligent Agents, IEEE Intl. Symposium on Intelligent Control [http://ieeexplore.ieee.org/search/searchresult.jsphistory=yes&queryText=%28%28action+evolution%29%3Cin%3Emetadata%29] *Intelligent agents