MetaVIM: Meta Variationally Intrinsic Motivated Reinforcement Learning for Decentralized Traffic Signal Control

TL;DR

MetaVIM addresses large-scale decentralized traffic signal control by framing it as a meta-learning problem where each intersection is a task and neighbor effects are captured through a learnable latent variable. The method combines a multi-head VAE (mVAE) with a policy network, using latent task information to make reward, transition, and policy functions shareable across tasks, while an intrinsic reward based on prediction biases stabilizes learning in dynamic multi-agent settings. Through CityFlow experiments across diverse real and synthetic road networks, MetaVIM achieves superior performance and transferability compared with conventional methods and multiple RL baselines, illustrating strong cross-scenario generalization and robustness to neighbor policies. The approach offers practical benefits for scalable deployment and rapid adaptation to new urban layouts without retraining from scratch.

Abstract

Traffic signal control aims to coordinate traffic signals across intersections to improve the traffic efficiency of a district or a city. Deep reinforcement learning (RL) has been applied to traffic signal control recently and demonstrated promising performance where each traffic signal is regarded as an agent. However, there are still several challenges that may limit its large-scale application in the real world. To make the policy learned from a training scenario generalizable to new unseen scenarios, a novel Meta Variationally Intrinsic Motivated (MetaVIM) RL method is proposed to learn the decentralized policy for each intersection that considers neighbor information in a latent way. Specifically, we formulate the policy learning as a meta-learning problem over a set of related tasks, where each task corresponds to traffic signal control at an intersection whose neighbors are regarded as the unobserved part of the state. Then, a learned latent variable is introduced to represent the task's specific information and is further brought into the policy for learning. In addition, to make the policy learning stable, a novel intrinsic reward is designed to encourage each agent's received rewards and observation transition to be predictable only conditioned on its own history. Extensive experiments conducted on CityFlow demonstrate that the proposed method substantially outperforms existing approaches and shows superior generalizability.

Figures (10)

• Figure 1: Illustration of the multi-intersection control scenario where policies and rewards are affected by neighboring policies, i.e., the instability of the multi-agent scenario.
• Figure 2: Illustration of incoming, outgoing lanes and 4-phase diagram.
• Figure 3: MetaVIM consists of a mVAE and a policy network. RL agent is augmented with a latent variable $m$, obtained by encoding over past trajectories $\tau_{ :t}$ to represent the task-specific information.
• Figure 4: Visualisation of the latent space. (a) and (b) denote the mean and variance of the latent variable respectively. The latent variable has 5 dimensions, each line is one latent dimension, the black line is the average.
• Figure 5: Illustration of intrinsic reward design.