Adaptive Traffic Lights with Monte Carlo Tree Search

Fixed-time and simple actuated signal plans adapt poorly to changing demand. Monte Carlo Tree Search (MCTS) offers online, simulation-based planning without an analytical traffic model.

A Python MCTS agent controls an isolated four-way intersection in SUMO 1.22 through TraCI. The state is the current signal phase plus the queue lengths on the approach lanes; actions select one of the two main green phases; the reward is the negative sum of halting vehicles over a short rollout. Each decision uses UCT selection, expansion, a rollout in which SUMO's state is saved, the phase is applied and the simulation is advanced, and backpropagation — after which SUMO's state is restored.

Deciding every 20 seconds, the controller switched phases in response to building queues — for example choosing the phase with a projected reward of −3.50 over −9.00 when one approach's queue grew. The paper is explicit about the prototype's limits (shallow, myopic search, a single intersection, a simple reward and no comparative baseline) and positions it as a foundation for scaled-up studies. Presented at MEKON 2025, Istanbul.