A large number of algorithms to generate behaviors of game agents have been developed in recent years. Most of them are based on artificial intelligence techniques that need a training stage. In this context, this paper proposes a minimal training strategy to develop autonomous virtual players using the NEAT neuroevolutionary algorithm to evolve an agent capable of playing the Flappy Bird game. NEAT was used to find the simplest neural network architecture that can perfectly play the game. The modeling of the scenarios and the fitness function were set to ensure adequate representation of the problem compared to the real game. The fitness function is a weighted average based on multiple scenarios and scenario-specific components. Coupling the minimal training strategy, a representative fitness and NEAT, the algorithm had a short convergence time (around 20 generations), with a low complexity network and achieved the perfect behavior in the game.
Training agents using Deep Reinforcement Learning methods is rapidly progressing in several fields and techniques like domain randomization have been demonstrated to improve the generalization ability of these agents. However, due to the black-box nature of the models, it is not easy to understand why an action was selected from a given input. Although prior research on Explainable Artificial Intelligence presents efforts to bridge this gap, is unclear what particular input features that contribute to a model’s generalizability. This work examines the main aspects that affect the behavior of game agents with varying robustness levels. By comparing specialized and generalized agents, we investigate what are the main differences and similarities present in these models when they select an action. To achieve this goal, we trained two agents with different robustness levels and applied Explainable Artificial Intelligence methods to highlight the key features on the input screen. We employed a mixed methods analysis, which provided important quantitative results on the agents’ performance as well as qualitative insights about their behavior. We are able to show that the visualization of generalized agents tends to be more interpretable since they concentrate on the game objects, whereas specialized agents are more spread along the whole input screen. This result constitutes an important step to understanding the behavior of game agents trained using Deep Reinforcement Learning with different training procedures.