We investigated how patients with Parkinson’s disease (PD) manage the explore–exploit trade-off in a structured reward-learning task. Patients were tested either on (N=34) or off (N=34) dopaminergic medication (levodopa), with age-matched polyneuropathy patients serving as controls (N=35). Behaviorally, patients off medication showed marked learning and decision-making deficits, characterized by overexploration and insufficient exploitation. To clarify the mechanisms underlying these impairments, we applied a computational model that combines similarity-based generalization with both random and uncertainty-directed exploration. The modeling results showed that impairments in patients off medication resulted from reduced generalization and increased uncertainty-directed exploration, but not greater random exploration. In contrast, exploration and generalization in patients on medication were comparable to the control group. Our findings highlight how dopamine depletion in PD impacts reward learning under uncertainty, suggesting a key role of dopamine in exploration and generalization.