Simulation of the evolution of autonomous adaptive agents

A model of evolving populations of self-learning agents is studied, and the interaction between learning and evolution is analyzed. Each agent is equipped with a neural network (NN) adaptive critic design for behavioral adaptation. The model is investigated by an example of a simple agent-broker that predicts stock price changes and uses these predictions for choosing strategies. Three variants of the model are compared, which include either learning and evolution, or only evolution, or only learning. It is shown that the model may exhibit the Baldwin effect; i.e., initially acquired adaptive policy of agents becomes inheritable during evolution. The behavior of the model agents is compared with the searching behavior of simple animals.