Abstract: The automated construction of informative heuristic functions based only on a domain description is a long-standing goal of automated planning. Recently, research has shown that deep reinforcement learning can be used to learn informative heuristic functions capable of solving puzzles such as the Rubik’s cube and sliding tile puzzles. In this talk, I will cover the preliminaries of deep reinforcement learning, advances in the application of deep learning and/or reinforcement learning to heuristic search, and examples of domains to which deep reinforcement learning has been successfully applied and domains for which further research is needed. Advances will include the DeepCubeA algorithm developed by myself and colleagues, generalizations of DeepCubeA to dynamically specified goals and domains with unknown transition functions, Q* search, DreamCoder, and Retro*. Applications will include puzzles, quantum computing, program synthesis, and chemical synthesis.

Bio: Forest Agostinelli is an assistant professor at the University of South Carolina. His research aims to use artificial intelligence to automate the discovery of new knowledge. He looks to apply his research to fields such as puzzle solving, chemical synthesis, program synthesis, theorem proving, robotics, and education. He led the creation of DeepCubeA, an artificial intelligence algorithm capable of solving puzzles such as the Rubik’s cube without human guidance. DeepCubeA has since been applied to problems in quantum computing, cryptography, and parking lot optimization. He earned his Ph.D. from the University of California, Irvine under the supervision of Professor Pierre Baldi. His homepage is located at https://cse.sc.edu/~foresta/.

Abstract: Multi-Robot Coverage Path Planning (MCPP) is a fundamental multi-robot coordination problem that aims to compute paths for multiple robots to jointly cover a given workspace. MCPP is critical in various practical applications, such as environmental monitoring, search and rescue operations, and agricultural automation. In this talk, I will explore the forefront of solving MCPP on graphs. First, I will examine a Mixed Integer Programming encoding for MCPP that optimally solves the Min-Max Rooted Tree Cover problem, showcasing a significant reduction in makespan (compared to existing MCPP methods) with a proven asymptotic suboptimality ratio. Next, I will introduce LS-MCPP, a method that bypasses traditional tree cover computations by employing a local search to significantly enhance efficiency and reduce makespan. Finally, I will present Multi-Robot Connected Fermat Spiral, an algorithmic framework that integrates computer graphics techniques with graph-based MCPP solutions to generate smooth, continuous coverage paths around irregular obstacles, which facilitates navigation in complex environments without the need for workspace decomposition.

Bio: Hang Ma is an Assistant Professor in Computing Science at Simon Fraser University. His research interests mainly lie in automated planning and multi-robot systems. His research has been recognized with several prestigious awards, including the ICAPS 2021 Best Dissertation and Best Demo, the AAAI 2021 New Faculty Highlights, an IFAAMAS 2020 Dissertation Award Runner-Up, and the ICAPS 2016 Best Robotics Paper.