Training of Agent Positioning Using Human’s Instruction

Abstract

In the real-world multiagent/multirobot problems, a position of each agent is an important factor to affect agents’ performance. In the real-world problem such as soccer, the agent(player)’s position should be changed based on the current environmental state. Because the real-world problem is generally dynamic and continuous, assigning the most desirable position for any state is not possible. We formalize this issue as a map from a focal point like a ball position in a soccer field to a desirable position of each agent. We conducted experiments showing that agent positioning are acquired efficiently through intuitive human operation and function approximation models using supervised learning, and conducted performance evaluation experiments to determine the most suitable model. In performance evaluation experiments, we evaluated the generalization capability for each model with datasets which several data are randomly removed from original dataset and dataset which several specific data are intentionally removed from original dataset. Experiment results showed that our proposed function approximation model combining Delaunay triangulation and linear interpolation produced the highest performance.