Inductive biases in theory-based reinforcement learning.

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Machine Learning

Background:

• Human learning in complex environments is a key cognitive science goal.
• Previous research often used simplified tasks, not reflecting real-world complexity.
• Video games offer a tractable proxy for studying complex, real-world learning.

Purpose of the Study:

• To investigate the inductive biases humans use in complex learning environments.
• To compare human learning biases with existing model-based reinforcement learning (RL) approaches.
• To enhance artificial intelligence learning by incorporating human-like semantic biases.

Main Methods:

• Utilized video games as experimental environments for complex learning.
• Developed a theory-based reinforcement learning (RL) framework.
• Identified and cataloged specific semantic biases used by human learners.
• Integrated these semantic biases into the theory-based RL system.

Main Results:

• Humans employ stronger inductive biases than syntactic constraints alone in theory formation.
• Identified a set of semantic biases that constrain the content of learned theories.
• Theory-based RL with added semantic biases demonstrated more human-like learning.

Conclusions:

• Human learning in complex environments relies on significant semantic biases beyond structural rules.
• Integrating these semantic biases into AI models can lead to more human-like artificial intelligence.
• This research bridges cognitive science and AI by modeling human learning biases.