Reinforcement learning plays a crucial role in optimizing decision-making under uncertainty, with the Multi-Armed Bandit (MAB) problem serving as a fundamental framework for balancing exploration and exploitation. However, classical approaches with fixed exploration rates often struggle to adapt to dynamic environments. This study evaluates the performance of key exploration strategies, including Epsilon-Greedy, Upper Confidence Bound (UCB), optimistic initialization, and gradient-based methods, within a controlled 10-armed bandit environment. By systematically comparing these methods across varying reward structures, the research highlights their strengths and weaknesses. The findings indicate that the Upper Confidence Bound (UCB) algorithm performs particularly well in dynamic environments due to its confidence-based exploration mechanism, achieving 5% higher average rewards and reaching optimal actions 20% faster than the Epsilon-Greedy algorithm set to ε = 0.1. While UCB offers adaptive precision, Epsilon-Greedy with a well-chosen parameter still provides a simple yet reliable trade-off between exploration and exploitation. Optimistic initialization promotes early exploration but may suffer from initial overestimation biases. The results contribute to the development of adaptive exploration strategies, providing valuable insights into decision-support systems in real-world applications. Future work will focus on hybrid approaches to further enhance algorithmic efficiency in non-stationary environments.