Human-level learning of complex novel tasks as theory-based modelling, exploration and planning.

Area of Science:

• Cognitive Science
• Artificial Intelligence
• Machine Learning

Background:

• Human learning is remarkably fast and versatile, unlike current machine reinforcement learning (RL) algorithms that require extensive data.
• Existing RL models struggle to replicate the efficiency and breadth of human task acquisition.

Purpose of the Study:

• To investigate human learning efficiency across diverse tasks.
• To propose a novel RL framework that accounts for rapid human learning.
• To develop an AI agent that mimics human learning capabilities.

Main Methods:

• Studied human learning on 90 video games to observe learning speed and strategies.
• Proposed a theory-based reinforcement learning (RL) approach using intuitive, causal theories.
• Developed an agent, EMPA (exploring, modelling, and planning agent), to instantiate the theory-based RL framework.

Main Results:

• Humans learn most tested games within minutes, demonstrating significant learning efficiency.
• EMPA successfully matched human learning efficiency and generalization capabilities.
• EMPA exhibited human-like exploration and learning dynamics across various game scenarios.

Conclusions:

• Theory-based RL, utilizing cognitively grounded intuitive theories, can effectively model rapid human learning.
• The EMPA agent provides a promising direction for developing more human-like AI learning systems.
• This research bridges the gap between artificial intelligence and cognitive science by modeling world knowledge and intuitive theories.