Humans combine value learning and hypothesis testing strategically in multi-dimensional probabilistic reward...

Area of Science:

• Cognitive Psychology
• Decision Science
• Machine Learning

Background:

• Complex decision tasks present numerous solutions, posing challenges for human learning.
• Traditional serial hypothesis testing may be inefficient with noisy feedback.
• Implicit reinforcement learning offers a parallel processing alternative.

Purpose of the Study:

• Investigate how individuals learn to solve complex, multi-dimensional problems.
• Model human learning strategies, comparing serial hypothesis testing and reinforcement learning.
• Examine the influence of task complexity on strategy selection.

Main Methods:

• Developed a multi-dimensional probabilistic active-learning task with configurable stimuli.
• Manipulated task complexity by varying relevant feature dimensions and information disclosure.
• Compared participant data against computational models of serial hypothesis testing and reinforcement learning.

Main Results:

• Evidence supports a hybrid learning strategy: hypothesis testing guided by reinforcement learning.
• Participants favored serial hypothesis testing in simpler tasks (fewer relevant dimensions).
• Participants relied more on parallel feature learning in complex tasks.

Conclusions:

• Humans strategically adapt their learning approach based on task complexity and available information.
• This adaptive strategy balances the efficiency of serial testing with the parallel processing of reinforcement learning.
• Findings illuminate the cognitive mechanisms underlying complex problem-solving and decision-making.