Active Inference — The Learn Arc, Part 48: Session §10.2 — Limitations

#activeinference #pomdp #ai #elixir

Series: The Learn Arc — 50 posts through the Active Inference workbench.
Previous: Part 47 — Session §10.1: Perception, action, learning

Hero line. Every framework earns trust by naming its own edges. Session 10.2 is the honest session — where Active Inference's assumptions break, where it handwaves, and which problems it is genuinely bad at.

Why limitations go in the book

A framework that only ever lists its successes is a marketing deck. A framework that names its failures is a tool. Session 10.2 is the latter — the session that keeps the other 47 honest.

Five beats

Policy enumeration does not scale. Eq 4.14 sums over policies. For short horizons and small action spaces this is cheap. For anything real — long horizons, large branching factors — it explodes combinatorially. Hierarchy helps, tree search helps, but the naive formulation does not survive contact with real robotics.
Continuous-time precisions are hard to choose. Chapter 8 was elegant; in practice, picking the right sensory/dynamical/parameter precisions is a black art. Miscalibrated precisions can make a continuous agent look deranged, and the framework does not give you a clean recipe.
Fitting is expensive. Chapter 9 is correct and clean, but the likelihood p(o, a | θ) is expensive to compute because it requires running the generative model for every candidate θ. MCMC runs that take hours on a laptop are standard.
The free-energy principle is not falsifiable as stated. As a normative claim ("an agent that minimises F exists"), it is close to a tautology. The constructive claim — "this specific agent minimises this specific F" — is falsifiable but small. The session is careful to distinguish the two; many critics conflate them.
Comparisons with RL are partial. Active Inference recovers many reinforcement-learning behaviors in the limit, and adds principled exploration. But on raw sample efficiency in well-specified reward tasks, deep-RL still wins. The framework's strengths are in calibrated uncertainty, transfer, and biological plausibility — not benchmark-bashing.

Why it matters

No serious researcher believes Active Inference is a finished product. Session 10.2 tells the reader exactly where the next decade of work is — planning at scale, precision calibration, fitting speed, cleaner empirical predictions. Knowing the limits is knowing the research agenda.

Quiz

Why does policy enumeration blow up so fast, and what is the canonical workaround?
Which part of the free-energy principle is (in its authors' view) more of a definition than a discovery?
If deep-RL wins on sample efficiency in simple tasks, where would you bet Active Inference wins instead?

Run it yourself

mix phx.server
# open http://localhost:4000/learn/session/10/s2_limitations

Cookbook recipe: limits/policy-horizon-blowup — plots wall-clock time of Eq 4.14 vs horizon depth for a fixed action space. The curve makes the combinatorial wall tangible.

Part 49: Session §10.3 — Where next. The capstone session. Which open problems are closest to tractable, which are bets, which are probably dead ends. The roadmap.

Powered by The ORCHESTRATE Active Inference Learning Workbench — Phoenix/LiveView on pure Jido.