Introduction
Model
Experimentation Future Works
● We focus only on D simulation league
● Strategy = Formations in Set-Play
● Basic method: Fedit
● Find a good strategy:
○ Difficult
○ Time consuming
○ Not always efficient
● Reuse existing strategies
● Similarity between teams
● General Specific
● Cluster already known teams
● Select automatically general suitable strategy
○ Team description
○ Description clustering
○ Find the most suited strategy for each cluster
● Strategy Selector
○ Estimation of corner kick’s success against clusters:
Bayesian Estimator
● Strategy Optimizer
○ Improve obtained results:
■ Use Earth Mover’s Distance
■ Observe light absorption parameter
■ Include PlayOn mode’s formations
● Strategy Optimizer V rd moduleW
● Optimization of an existing corner kick formation
● Implementation of Firefly Algorithm
● Start from a handmade formation
● 8 generations of 5 formations
● Computation time: days
● Work distributed on: machines Generations
Fitness
● Optimize general strategy for any particular team
○ Select the strategy associated to the
corresponding cluster
○ Optimize it by
heuristic/genetic methods
● For any new team
○ Classify the team
○ Optimize the general strategy
● We focus only on Strategy Selector and Strategy
Optimizer
● No relevant results at the moment
Selecting the Best Strategy Against a Particular Opponent Team
in RoboCup D Simulation
Henrio Jordan, Henn Thomas, Nakashima Tomoharu, Akiyama Hidehisa, Mifune Satoshi
Proposed Model
Fedit2
Experimentation Result
