A multiagent framework for learning dynamic traffic management strategies

There is strong commercial interest in the use of large scale automated transport robots in industrial settings (e.g. warehouse robots) and we are beginning to see new applications extending these systems into our urban environments in the form of autonomous cars and package delivery drones. This new technology comes with new risks—increasing traffic congestion and concerns over safety; it also comes with new opportunities—massively distributed information and communication systems. In this paper, we present a method that leverages the distributed nature of the autonomous traffic to provide improved traffic throughput while maintaining strict capacity constraints across the network. Our proposed multiagent-based dynamic traffic management strategy borrows concepts from both air traffic control and highway metering lights. We introduce controller agents whose actions are to adjust the robots’ perceived “costs” of traveling across different parts of the traffic network. This approach allows each robot the flexibility of using its own (potentially proprietary) navigation algorithm, while still being bound by the “rules of the road.” The control policies of the agents are defined as neural networks whose weights are learned via cooperative coevolution across the entire traffic management team. Results in a real world road network and a simulated warehouse domain demonstrate that our multiagent traffic management system provides substantial improvements to overall traffic throughput in terms of number of successful trips in a fixed amount of time, as well as faster average traversal times.