Active Inference — The Limitations

This article discusses the limitations of the Active Inference framework, including policy enumeration scaling issues, difficulty in choosing continuous-time precisions, expensive model fitting, the non-falsifiability of the free-energy principle, and the partial comparisons with reinforcement learning.

Details

The article discusses five key limitations of the Active Inference framework. First, the policy enumeration approach described in Eq 4.14 becomes computationally intractable for real-world problems with long horizons and large action spaces, even though techniques like hierarchy and tree search can help. Second, choosing the right continuous-time precisions for sensory, dynamical, and parameter components is a black art, and can lead to unstable agent behavior if miscalibrated. Third, the likelihood computation required for model fitting, as described in Chapter 9, is expensive and can take hours on a laptop using MCMC methods. Fourth, the free-energy principle is not easily falsifiable as stated - the normative claim is close to a tautology, while the constructive claim is more limited. Finally, while Active Inference recovers many reinforcement learning behaviors and adds principled exploration, it still lags behind deep RL in terms of raw sample efficiency on well-specified reward tasks, though it has strengths in calibrated uncertainty, transfer learning, and biological plausibility.