Diversity in multi-agent systems for successful real-world deployments

点击翻译按钮获取中文摘要

Machine learning, a data-driven type of artificial intelligence, has revolutionized multiple industries. Not only does this allow for higher performance but it can also enable previously unrealizable opportunities. Unfortunately, learning can be impractically slow, particularly when data is expensive or slow to gather. For example, rather than letting a robot learn for days in the real world, which costs both time and money, why not first learn in simulation and then transfer to the real world? Unfortunately, because simulations are never perfect, some type of adaptation and simulation change strategies are needed, typically called Sim2Real. While common in single-agent settings, Multi-agent Sim2Real (MAS2R), has not been studied, even though it is a critical component to many real-world potential breakthroughs. This project will extend and test MAS2R, considering not only when situations change but also when different types of actors are in the environment. MAS2R will be applied to high-fidelity simulations in:Autonomous DrivingSmart GridSmart BuildingsCommercialization options will be explored, with one or more paths forward towards commercialization decided on for next steps to help Alberta companies, post-award.