Method for selecting tactical strategies for game artificial intelligence agents using topsis

Abstract

The paper addresses the problem of tactical strategy selection by game artificial intelligence (AI) agents in two-team combat scenarios. A combat environment model is proposed in which the sets of allies and enemies for each agent are determined dynamically within the detection radius, while the agent’s current state is evaluated according to six criteria: ammunition, health, the number of nearby allies and enemies, the availability of cover, and potential threat. A tactical strategy selection method based on TOPSIS is developed. In the proposed method, the agent’s state indicators are transformed into suitability scores for the strategies attack, defense, and retreat, taking into account the direction of influence of each criterion on each strategy; the resulting alternatives are then ranked. A numerical example is provided, together with a sensitivity analysis of strategy selection with respect to the number of opponents. The method was implemented and evaluated in a computer simulation of combat between two teams of five agents using a paired design with side switching over four hundred battles. The team controlled by the proposed method won sixty-seven percent of the battles, whereas the team whose agents always selected the attack strategy won four percent; twenty-nine percent of the battles ended in a draw. The ablation study showed that random strategy selection under the same movement rules provides no advantage, whereas TOPSIS-based aggregation is significantly more effective than the weighted-sum approach applied to the same decision matrix, yielding sixty-seven percent versus fifty-six and a half percent wins. It was found that, due to the compensatory nature of the aggregation procedure, an agent may select the attack strategy even when no ammunition is available. Introducing a non-compensatory admissibility constraint for this strategy increased the win rate to eighty percent and reduced the draw rate to sixteen percent. The advantage of the method is preserved under variations in the criterion weights and environment parameters, while its magnitude depends most strongly on the protective effect of cover.