Action evolution for intelligent agents
Tessendorf, Rodney E.
Van Schoor, George
MetadataShow full item record
This paper introduces an alternative method of reinforcement learning for intelligent agents. The aim with this method is to closer emulate the natural thought processes of problem solving. This paper only provides a conceptual description of this method. However, it puts forward realistic implementations with exciting new implications, using proven techniques. After a brief discussion of reinforcement learning, the newly suggested method is described. Reinforcement learning methods can be divided according to two main strategies. The first strategy uses evolutionary methods while the second makes use of incremental back-propagation methods. An overview is given of the methods used for implementing the popular actor-critic architecture in these strategies. This leads to a generic architecture for intelligent controllers. The alternative method of action reinforcement is then demonstrated at the hand of this architecture. This representation of the controller emphasizes some cognitive attributes and encourages the development of advanced intelligent controllers