Countering adversarial strategies in multi-agent virtual scenarios

Mutual modeling and plotting between virtual characters is an important issue in scenarios with an interactive multi-agent drama and games. This paper provides a framework for agents to model the intentions of others and to select the optimal strategy by simulating multi-level mutual modeling. The resulting plan considers multiple rounds of counter-strategy and possible adversarial reactions, creating deeper strategic interaction. The framework use a hierarchical tree similar to an AND-OR tree in structure to represent both the plans of self and of the adversary, allowing agents to reason about multi-level countering strategies and decide whether a strategy is optimal or worth taking. The method for countering one-level strategy may work if an agent's opponent is completely unaware of my intention to counter it. If the opponent is clever, however, it will predict the agent's intention to counter his plan and will try to devise schemes to obstruct it. Then the agent will take action to prevent its counterplan from being countered. It is evident that a "clever" agent should be capable of simulating multi-level mutual modeling in mind. Finally, the framework for multilevel modeling should be incorporated in the architecture of the virtual agent, who has a sensor to receive events in the virtual world and an actuator to impose actions. These action results cause the structure of and probabilities in the intention trees to be updated. A counterplanning agent should dynamically adjust its strategies as a response to the adversarial actions. There are still some issues to be dealt with. Since an action could fail disastrously rather than "just fails", the agent should expect the possibility of failure and prepare remedies. Time control in executing countering actions is also an important issue. Still, we hope this work brings a step closer to automatic scenario performance.