02 / Selected work

Multi-Agent Traffic Signal Control

Multi-agent reinforcement learning

Four traffic signals learning to cooperate. Each intersection runs its own Double DQN agent and shares queue information with its neighbours, so the grid balances load instead of four controllers fighting each other.

average wait-time reduction

94.3%average wait-time reduction

better per-intersection on a 4-agent network

68%better per-intersection on a 4-agent network

over the single-agent baseline

86.8%over the single-agent baseline

Team project