Short-Term Memory Mechanisms in the Goal-Directed Behavior of the Neural Network Agents

Modern machine learning methods are not able to achieve level of adaptability comparable with one that observed in the animals’ behavior in complex environments with numerous goals. This fact necessitates the investigation of general principles for the formation of complex control systems able to provide effective goal-directed behavior. We have developed original neuroevolutionary model for the agents situated in stochastic environments with hierarchy of goals. The paper provides the analysis of the evolutionary dynamics of agents’ behavioral strategies. Analysis’s results demonstrate that evolution results in neural network controllers that allow agents to store information in short-term memory via several neurodynamical mechanisms and use it for behavior based on alternative actions. During the study of neuronal basics of the agents’ behavior we found that neurons’ groups could be responsible for different stages of behavior.