One Agent per Zone: Decentralised Active Inference under Partial Observability for Multi-Zone HVAC Control on Edge Hardware
Decentralised HVAC control in multi-zone buildings is difficult because each zone controller sees only local state while the zones share a single air loop. This paper assigns every zone an independent active-inference agent that learns its own transition model online, infers hidden coupling via a latent disturbance factor, and adjusts its energy weight in response to a live congestion signal drawn from the metered power of other floors, with no inter-agent messages. On five noisy test years the fleet achieves the lowest mean building peak ($88.3~\mathrm{kW}$) and the best coincidence factor ($0.929$) among nine controllers, while running the full fleet on a Raspberry Pi 5 with a p95 control latency of $442~\mathrm{ms}$.
The paper makes a clear empirical case that independent active-inference agents can stagger HVAC demand across floors without direct communication, achieving the lowest mean annual peak ($88.3~\mathrm{kW}$) and best coincidence factor ($0.929$) on five noisy test years. However, the economic advantage is fragile: under the evaluated tariff the fleet actually costs $1.6$k more per year than a simple reactive rule because its $10\%$ energy penalty outweighs the demand savings, and it only breaks even at a demand charge of about $75/\mathrm{kW}$-month. Comfort is also weaker than the proactive baseline, and degrades severely in an unseen hot climate ($\mathrm{CS}_{occ}=0.798$), so the practical payoff depends heavily on tariff structure and weather.
The ablation study in Table V is the strongest element: it cleanly isolates that neither the congestion signal nor adaptive weights alone reduce the peak, but their combination does, and it shows that delayed or constant signals fail, ruling out trivial explanations. The authors also provide an unusually transparent baseline by sweeping the proactive RBC's cooling margin, reporting deadband-violation counts, and presenting year-by-year paired differences. The edge-deployment measurement is concrete, with the full MQTT round-trip on a Raspberry Pi 5 reported at $442~\mathrm{ms}$ p95.
The foremost limitation is economic rather than algorithmic: the AIF fleet consumes about $10\%$ more energy than the reactive RBC and, under the $26/\mathrm{kW}$-month tariff evaluated, ends up more expensive despite lower peaks. Comfort is another weakness---the fleet is the least comfortable of the strong controllers on the test years, and in an unseen hot climate its occupied comfort score collapses to $\mathrm{CS}_{occ}=0.798$. Generalisability is narrow: a single ASHRAE prototype building, one tariff, deterministic occupancy, and only one training run per MADDPG variant. The claim that agents exchange no messages is technically true at the control layer, but they all depend on the same centrally metered floor-power signals, so independence is bounded by shared infrastructure.
The quantitative evidence supports the peak-shaving and coincidence claims: Table I shows the AIF fleet beats all baselines on annual peak and CF across five noisy years, and Table II confirms the advantage is consistent year-by-year against most baselines. The authors fairly note that the advantage over the proactive RBC is not robust (only $3/5$ years). They also honestly concede that the same live meters given to MADDPG yield similar staggering, so the live load information---not active inference alone---is what enables the staggering. Where the evidence is thinner is in cost and comfort: the tariff analysis reveals the AIF fleet is not Pareto-optimal---it loses to the proactive RBC on comfort and to the reactive RBC on cost under current pricing.
Reproducibility is mixed. On the positive side, all code, trained models and benchmark harnesses are released under Apache 2.0, and the simulation uses a standard Sinergym/EnergyPlus ASHRAE 90.1 OfficeMedium model. On the negative side, the MADDPG baseline was trained only once with no cross-seed evaluation, so its error bars in Table I reflect weather variation rather than training variance. The five noisy weather seeds are not explicitly listed, and the $50$ simulated years of MADDPG training ($\sim 3.5$ million updates) create a compute barrier. The AIF agents themselves require five years of online adaptation before freezing, which is feasible but time-consuming to replicate.
Multi-zone HVAC control is a shared-resource problem: zones on a floor share one fan and cooling coil, so each zone's demand affects its neighbours and the building's peak power. One controller per zone scales well and isolates faults, but each controller then observes only its own zone, not the shared air loop. In this work, we present a decentralised controller that assigns each zone an independent active-inference agent, which plans by minimising expected free energy and learns its zone dynamics online. The agents exchange no messages, so none depends on peers or a coordinator; each accounts for the unobserved coupling through a latent disturbance and a congestion signal derived from the metered power of the other floors. Fifteen agents are evaluated in EnergyPlus on a Sinergym office building against three rule-based controllers and three MADDPG fleets, one given the same floor meters. On five noisy weather years unseen during training, the fleet attains the lowest mean building peak (88.3 kW, 3 to 17 % below the baselines) and staggers the floor peaks better than any baseline, with a coincidence factor (building peak over the sum of the floor peaks) of 0.929 versus 0.936 to 0.994. Compared with a reactive rule at similar comfort (0.912 versus 0.916), it lowers the sum of monthly peaks, on which demand charges are billed, by 33 kW (4 %) in every year, at the cost of about 10 % more energy. It costs 9% less per year than the proactive rule (at lower comfort) and 1% less than the MADDPG fleet given the same floor meters, in every test year. The whole fleet runs on a single low-cost Raspberry Pi 5, closing the observation-to-action loop over MQTT in 442 ms at p95, far below the 15-minute control step.
Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.
No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.