Traffic controllers are designed to use sensing technology and control policies to optimize vehicle movements on transportation networks. However, these controllers are typically hard to operate or train to act as a collectively intelligent system. In addition, using artificial intelligence (AI) to train control agents is challenging due to the lack of training data needed for imitated learning methods. In this work, we use the musicians' sense of improvisation to train agents that can make optimal control actions based on traffic state input. We present the training method and results of the Green Light SONATA (Signal Operation with Neuro-fuzzy Acoustic Tuning Application) agents using Markov state estimation and value iteration algorithms. Ultimately, we show the feasibility of this method and how it can be used to assemble a system of SONATA agents to work in a collectively intelligent multi-agent system.
Back to Workshop IV: Modeling Multi-Scale Collective Intelligences
Copyright. All Rights Reserved.