Punctuated anytime learning for evolving a team
Gary B. Parker, Harvey J. Blumenthal · 2003
Learning heterogeneous behaviors for robots to cooperate in the performance of a task is a difficult problem. Evolving the separate team members in a single chromosome limits the capacity of the genetic algorithm to learn. Evolving the separate team members in separate populations promotes specialization and gives the genetic algorithm more flexibility to produce a solution, but can be either computationally prohibitive or result in credit assignment complications. In this paper, we apply punctuated anytime learning to assist in the co-evolution of separate team member populations. A box-pushing task is used to show the success of this method.