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

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Computational Neuroscience

Background:

• Decision-making flexibility is crucial for human cognition and adapting to changing environments.
• Artificial intelligence (AI) agents, particularly those using deep neural networks, are used to simulate human cognitive processes.
• Both supervised learning (SL) and reinforcement learning (RL) are employed to train AI agents, but their efficacy in replicating human decision-making flexibility is not fully understood.

Purpose of the Study:

• To compare the effectiveness of supervised learning (SL) and reinforcement learning (RL) in training AI agents with human-like decision-making flexibility.
• To investigate how different learning paradigms influence an agent's ability to adapt strategies under varying decision criteria.

Main Methods:

• Identical deep artificial neural network architectures were trained using both SL and RL paradigms.
• Agents were tasked with a memory-based decision task under three distinct criteria: precise, conservative, and liberal.
• Performance was evaluated based on the agents' ability to adapt their decision-making strategies.

Main Results:

• Agents trained with both SL and RL performed accurately under the precise decision criterion.
• Only RL-trained agents successfully adapted to the conservative and liberal decision criteria, demonstrating superior flexibility.
• RL-trained agents significantly outperformed SL-trained agents in adapting to diverse decision-making scenarios.

Conclusions:

• Reinforcement learning (RL) is a more effective paradigm than supervised learning (SL) for modeling human-like decision-making flexibility in AI agents.
• These findings offer valuable insights for developing AI systems that better replicate human cognitive functions for real-world applications.