next up previous
Next: Kicking Up: The CMUnited-98 Champion Simulator Previous: Updating the World Model

Agent Skills

 

Once the agent has determined the server's world state for cycle t as accurately as possible, it can choose and send an action to be executed at the end of the cycle. In so doing, it must choose its local goal within the team's overall strategy. It can then choose from among several low-level skills which provide it with basic capabilities. The output of the skills are primitive movement commands.

The skills available to CMUnited-98 players include kicking, dribbling, ball interception, goaltending, defending, and clearing. The implementation details of these skills are described in this section.

The common thread among these skills is that they are all predictive, locally optimal skills (PLOS). They take into account predicted world models as well as predicted effects of future actions in order to determine the optimal primitive action from a local perspective, both in time and in space.

One simple example of PLOS is each individual agent's stamina management. The server models stamina as having a replenishable and a non-replenishable component. Each is only decremented when the current stamina goes below a fixed threshold. Each player monitors its own stamina level to make sure that it never uses up any of the non-replenishable component of its stamina. No matter how fast it should move according to the behavior the player is executing, it slows down its movement to keep itself from getting too tired. While such behavior might not be optimal in the context of the team's goal, it is locally optimal considering the agent's current tired state.

Even though the skills are predictive, the agent commits to only one action during each cycle. When the time comes to act again, the situation is completely reevaluated. If the world is close to the anticipated configuration, then the agent will act similarly to the way it predicted on previous cycles. However, if the world is significantly different, the agent will arrive at a new sequence of actions rather than being committed to a previous plan. Again, it will only execute the first step in the new sequence.





next up previous
Next: Kicking Up: The CMUnited-98 Champion Simulator Previous: Updating the World Model



Peter Stone
Mon Nov 30 20:08:29 EST 1998