Agents Trained through Reinforcement Learning Exhibit Human-Like Decision-Making Flexibility

Decision-making flexibility is a fundamental aspect of human cognition, allowing individuals to adapt their strategies based on changing situations. To better understand human cognitive processes, researchers have developed artificial intelligence (AI) agents that simulate human behavior and neural activity. Recent research shows that agents with deep artificial neural networks as their core can simulate human behavior through either supervised learning (SL) or reinforcement learning (RL). However, it remains unclear which learning paradigm is more effective at creating agents with human-like decision-making flexibility. In this study, we trained agents with identical architectures using both SL and RL paradigms to complete a memory-based decision task under three distinct decision criteria. Under the precise decision criterion, agents trained with both learning paradigms were able to make accurate decisions. However, under the conservative and liberal decision criteria, only RL-trained agents successfully mastered the task. These results demonstrate that RL-trained agents outperformed their SL-trained counterparts, exhibiting greater adaptability to diverse decision criteria. Our findings highlight RL as a more effective paradigm for modeling human-like decision-making flexibility. This provides valuable insights for developing AI systems that more accurately replicate human cognitive functions in real-world scenarios.