Multi-task reinforcement learning in humans

Area of Science:

• Cognitive Science
• Neuroscience
• Artificial Intelligence

Background:

• Human intelligence excels at transferring knowledge across tasks and generalizing to new ones.
• Multitask reinforcement learning in humans remains under-investigated.
• Understanding how humans learn and adapt in dynamic environments is crucial.

Purpose of the Study:

• To investigate human behavior in multitask reinforcement learning.
• To compare human strategies with computational models.
• To elucidate the mechanisms of knowledge transfer in complex decision-making.

Main Methods:

• Participants engaged in a two-step decision-making task with varying features and reward functions.
• Human behavior was compared against two multitask reinforcement learning algorithms (policy mapping and value function approximation) and standard algorithms.
• Data from three exploratory and one large confirmatory experiment were analyzed.

Main Results:

• Participants capable of learning the task demonstrated a strategy involving mapping previously acquired policies to novel scenarios.
• Human performance aligned with algorithms employing policy mapping for knowledge transfer.
• The findings suggest a specific cognitive strategy for navigating complex, changing environments.

Conclusions:

• Humans utilize policy mapping to generalize learning in multitask reinforcement learning scenarios.
• This cognitive strategy is comparable to advanced artificial intelligence approaches.
• The study enhances understanding of human adaptability and learning in dynamic, complex environments.