Active Inference Enhances Traffic Control in Noisy IoT Environments

Researchers have introduced an active inference controller aimed at improving adaptive traffic signal management, particularly in urban environments characterized by noisy and nonstationary IoT data. Traditional traffic controllers often degrade under conditions like sensor occlusion, adverse weather, or unpredictable demand, and even learned policies can lack transparency. This new approach addresses these issues by dynamically selecting traffic phases through the minimization of expected free energy, based on Gaussian beliefs about congestion levels. The proposed controller provides a fully traceable decision pipeline, which is a significant advantage for auditing and understanding its behavior. It was rigorously benchmarked in a SUMO traffic simulator against both a rule-based heuristic and a deep Q-network (DQN) across various scenarios that progressively introduced more noise and nonstationarity, including sensor failures and stochastic accidents. Results indicated that the active inference controller achieved the lowest idle times and CO2 emissions in the most challenging, noisy scenarios, outperforming the DQN. While there was a modest trade-off in bus priority service rate and phase switch frequency, the overall gains in efficiency and environmental impact in adverse conditions highlight the potential of this physics-informed approach for robust urban infrastructure management.