Generating Controller of GR(1) Synthesis and Reinforcement Learning in Game-Solving

다국어 초록 (Multilingual Abstract)

This paper presents a comparative study of GR(1) Synthesis and Reinforcement Learning in the context of controller generation for the Moving Obstacle Evasion Problem. GR(1) Synthesis, a formal method in computer science, provides a systematic approach to automatically generate a controller that satisfies a given set of logical specifications. On the other hand, Reinforcement Learning, a type of machine learning, approximates the optimal solution by learning from the environment and updating its strategy based on a loss function. While GR(1) Synthesis guarantees an optimal solution, it may fail to generate a controller in certain configurations. In contrast, Reinforcement Learning can provide a suboptimal but feasible solution in all configurations. The results demonstrate the potential of Reinforcement Learning as a viable alternative for controller generation in scenarios where GR(1) Synthesis is unrealizable. This study contributes to the understanding of the strengths and limitations of both methods and provides insights for their application in system design.