Traffic simulation showcasing a traffic signal control agent, modeled as a deep artificial Q network, trained using reinforcement learning.
First 30 seconds the agent has learned nothing, and is taking exploratory actions 100% of the time, large queue accumulates, poor performance. The last 30 seconds, after significant learning, the agent is exploiting 100% of the time, taking actions which yield high reward (reduction in the number of queued vehicles) very small queues forms, high performance.
Simulation training time equivalent to 104 days, Q-learning, e-greedy policy, 4 layer (3 convolutional + 1 dense) architecture. SUMO simulation software, Theano + Keras code for ANN.
Post time: Jun-12-2017