Agent-space transfer learning for sign-based world model

Abstract

Several algorithms were studied for transfer learning problem for a foraging task. These algorithms are suitable for a semiotic control system to facilitate experience transfer between two learning agents with different agentspace descriptions of the state in a predicate form. They require that the target agent’s description of the task is a subset of the source agent’s description. The algorithms are based on the Q-learning algorithm and uses a custom transfer function to initialize the target agent’s state-action table, which matches the corresponding predicates. The test problem used to test the algorithms is a foraging task, where an agent must gather randomly generated food in a grid world. The results show that they provide improvement in the learning curve for the target agent after transferring experience from an agent with more input predicates. The improvement is inconsistent and sometimes does not bring noticeable difference in performance, but the target agent performs at least as good as an agent with no prior experience.