Branching time active inference: Empirical study and complexity class analysis.

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Reinforcement learning

Background:

• Active inference is a powerful brain modeling framework, but suffers from computational complexity in policy computation.
• Existing solutions like Monte Carlo Tree Search and structure learning have shown promise.
• Branching-time active inference (BTAI) combines these ideas but requires empirical validation.

Purpose of the Study:

• To empirically evaluate the performance of branching-time active inference (BTAI).
• To assess BTAI's effectiveness in solving complex tasks, particularly in comparison to standard active inference (AcI).
• To analyze BTAI's scalability and efficiency against established algorithms.

Main Methods:

• Experimental study of BTAI using a maze-solving agent.
• Comparison of BTAI with standard active inference (AcI) on graph navigation tasks of varying sizes.
• Benchmarking BTAI against POMCP on the frozen lake environment.
• Comparative analysis with Fountas et al. (2020) on the dSprites dataset.

Main Results:

• Improved prior preferences and deeper search in BTAI mitigate local minima issues.
• BTAI scales to larger graphs where AcI becomes intractable due to exponential complexity.
• BTAI and POMCP achieve comparable rewards on the frozen lake environment, with specific conditions favoring each.
• BTAI demonstrates competitive performance against Fountas et al.'s approach on the dSprites dataset.

Conclusions:

• BTAI provides a computationally efficient and scalable alternative to standard active inference.
• The empirical validation confirms BTAI's ability to handle complex problems intractable for other methods.
• BTAI represents a significant advancement in applying active inference to complex decision-making tasks.