Accurate reinforcement learning is a challenge in multi-agent reinforcement learning, which relies heavily on mathematical model and algorithm development. When applied to the research challenges of multi-agent reinforcement learning, traditional genetic algorithms provide unsatisfactory results and are unable to resolve the algorithm research issues. As a result, this study examines previous work on multi-agent reinforcement learning and suggests a model and method for the field that relies on stochastic countermeasures. In order to minimize interference in mathematical model and algorithm research, the strategy gradient theory is used to identify the components that have an impact. The indicators are then classified based on the needs of the study. Next, a mathematical model and algorithm research scheme is developed for stochastic countermeasures using the strategy gradient theory. The outcomes of this study are then thoroughly examined. In terms of mathematical model and algorithm research time, mathematical model and algorithm research impact factor accuracy, and other assessment criteria, the MATLAB simulation results demonstrate that the random countermeasure outperforms the classic genetic algorithm.