Starting Position Bias in Multiplayer Games Through Reinforcement Learning
Abstract
Ensuring fairness in multiplayer games is one of the most important elements. To achieve fairness, the game needs to be well-balanced, and this has become a serious challenge for game developers in recent times. In this study, a multi agent PPO based testing method was developed to automatically detect and reduce positional bias in games. The study uses the Dungeon Escape environment of the ML-Agents library in Unity 3D, and this environment has been modified to suit individual competition for the purpose of the test. In this environment, consisting of a key, a dragon, and a door, the individual goal of the agents is to win by taking the key and escaping through the door. To measure the effect of the agents' starting positions on the game outcome, three different starting areas (Close, Near, Far) were defined according to their distance from the door. Identical agents were trained using the Proximal Policy Optimization (PPO) algorithm under a shared policy architecture. Two different test configurations were created to evaluate positional inequality: First, 10,000 episodes were run in a symmetrical arrangement with three identical agents positionally permuted; second, 10,000 episodes were run in an asymmetrical arrangement with six agents (1 Close, 2 Near, 3 Far) distributed to numerically compensate for disadvantaged positions. The test results show that spatial proximity to the door positively affects the winning rate (p < 0.0001). Furthermore, having a numerical advantage in terms of number of agents does not reduce individual positional risk, and the average reward amount masks this positional bias in some cases. Our findings demonstrate that initial positional bias is a purely structural consequence of the environment design, independent of the training process, and we offer to utilize this method as a stable testing tool for the early detection and resolution of inequalities in games.