Evolution of Goal-Directed Behavior from Limited Information in a Complex Environment

In this paper, we apply an evolutionary algorithm to learning behavior on a novel, interesting task to explore the general issue of learning e ective behaviors in a complex environment that provides only limited perception and goal-feedback. Our speci c approach evolves behavior in the form Articial Neural Networks with recurrent connections. We apply our approach to learn e ective behavior for a non-standard mazenavigation problem that is characterized by aspects of problems that are di cult to approach via other methods. Di cult aspects of the speci ed problem include the inability to sense all task-relevant state at any given time the problem of hidden state", and limited feedback with respect to success or failure. We observe evolved networks to perform very well on the target problem. Further ndings include adaptation to noise in action selection, performance proportional to memorycapacity, and improved performance when network weights are transferred from training on one ...

• Type: Conference
• Year: 1999
• Where: GECCO
• Authors: Matthew R. Glickman, Katia P. Sycara
• Keywords: Approach Evolves Behavior, Di Cult, Ective Behavior, GECCO 1999, Optimization