Agent Architecture   

Home
Members
Overview
Research Areas
Download
Competitions
Sponsors
News
Links

Agent Architecture and High Level Control Flow

This outlines our agent architecture and the control flow of our high-level decision module. FC Portugal, team strategy includes the knowledge about tactics, formations, player types, situations, intelligent communication and perception strategies and opponent modeling strategies. These structures are instantiated initially by a human coach according to the opponent team characteristics.

The main control loop of our agents uses perception interpretation and action prediction to update the world state and then uses a high-level decision module to decide the next action. Information model in FC Portugal is a multi-level structure with data at four levels of abstraction:

• ·     Global Situation Information – High-level information like result, time, game statistics (shoots, successful passes, possession, etc.) and opponent behavior, necessary to decide the team tactic at a given moment;

• ·     Situation Information – Information relevant to the selection of the appropriate formation and for the SBSP, SLM and ADVCOM mechanisms;

• ·     Action Selection Information – Set of high-level parameters used to identify active situations and to select appropriate ball possession or ball recovery behaviors;

• ·     World State – Low-level information, including positions and velocities of the players and of the ball.

The high-level decision module decides not only the player action but also the tactic, formation, positioning and player type of the player at a given moment. This high-level decision module starts by initializing some internal structures and entering the agent main control loop. This loop starts by running the STRATEGY module in which the agent uses game global information to decide the current tactic and situation information to decide the formation. Then, the agent analyses Dynamic Positioning and Role Exchange possibilities, performs intelligent communication analysis and runs the strategic looking module.

If the game is stopped, stopped game positioning is performed. If an active situation is not identified, meaning that the agent does not need to get immediately into active behavior, SBSP is performed. If the situation is active then, if the agent has the control of the ball, an appropriate ball possession action is selected and if not, an appropriate ball recovery action is selected. The loop closes with modules not directly connected with decision-making: action execution and multi-level world state update.

Back to Top

Back to FC Portugal Approach Overview Page