Active Inference AI beats DQN in noisy traffic signal control
New active inference controller cuts idle time by 20% in harsh IoT conditions
Urban traffic signals struggle with sensor noise, weather, and fluctuating demand. A team from multiple institutions introduces an active inference controller that selects phases by minimizing expected free energy (EFE) over Gaussian beliefs about per-direction congestion. This yields a fully traceable decision pipeline, unlike black-box deep Q-networks (DQN).
Tested in SUMO across four scenarios (including sensor occlusion, adverse weather, and stochastic accidents), the active inference controller consistently outperformed a rule-based heuristic and DQN. In the noisiest scenarios, it achieved 56,977 seconds of idle time and 29.12 kg of CO2 emissions, compared to DQN's 71,741 seconds and 30.56 kgāa 21% reduction in idle time and 5% lower emissions. The trade-off: slightly lower bus priority service rate and more frequent phase switches. The paper is submitted to IEEE WF-IoT 2026.
- Active inference controller uses expected free energy minimization over Gaussian beliefs for fully traceable decisions
- In SUMO simulations, it achieved 56,977s idle time and 29.12kg CO2 vs DQN's 71,741s and 30.56kg under noisy, nonstationary conditions
- Method tested across four scenarios including sensor occlusion, adverse weather, and random accidents with 100 evaluations each
Why It Matters
Auditable, robust AI for smart city traffic control could cut congestion and emissions without sacrificing transparency.