The iterated prisoner’s dilemma (IPD) game serves as a crucial framework for exploring cooperation and betrayal, yet it presents several challenges. These challenges include the dominance of selfish strategies in noisy environments, the difficulty fixed strategies face in consistently outperforming all opponents, the risk of overfitting in uniform experimental setups, and the limited exploration of multi-objective optimization. In this study, we employed a genetic algorithm (GA) to evaluate four strategy representations across six evolutionary systems with multiple objectives. Strategies were trained and tested under varying noise conditions using round-robin tournaments. Our findings revealed that, across different evolutionary systems and noise levels, selfish strategies yielded higher rewards in single-objective scenarios, while cooperative strategies excelled in multi-objective settings. Furthermore, both linear and neural network representations demonstrated robust performance. An analysis of feature weights and decision bits indicated that cooperative strategies adapt their behaviors based on opponent actions and the progression of rounds, thereby enhancing their robustness. Additionally, incorporating diverse noise levels into the training process further strengthens the adaptability of these cooperative strategies, enabling them to effectively counter new opponents.