Using ontology to guide reinforcement learning agents in unseen situations

摘要

In multi-agent systems, goal achievement is challenging when agents operate in ever-changing environments and face unseen situations, where not all the goals are known or predefined. In such cases, agents need to identify the changes and adapt their behaviour, by evolving their goals or even generating new goals to address the emerging requirements. Learning and practical reasoning techniques have been used to enable agents with limited knowledge to adapt to new circumstances. However, they depend on the availability of large amounts of data, require long exploration periods, and cannot help agents to set new goals. Furthermore, the accuracy of agents’ actions is improved by introducing added intelligence through integrating conceptual features extracted from ontologies. However, the concerns related to taking suitable actions when unseen situations occur are not addressed. This paper proposes a new Automatic Goal Generation Model (AGGM) that enables agents to create new goals to handle unseen situations and to adapt to their ever-changing environment on a real-time basis. AGGM is compared to Q-learning, SARSA, and Deep Q Network in a Traffic Signal Control System case study. The results show that AGGM outperforms the baseline algorithms in unseen situations while handling the seen situations as well as the baseline algorithms.