Small farms are thought to produce around a third of the global crop supply. But, in the wake of the climate crisis, their existence is increasingly vulnerable to changes in the spatial and temporal availability of water. The small-scale irrigation systems which water these farms present a social-ecological dilemma: Upstream farms have prevailing access to a canal’s resources, but all farms along the canal must contribute to maintaining the irrigation infrastructure. Thus, it is key to assess the social mechanisms which promote resilience in these systems and, more widely, in complex social-ecological dilemmas under changing conditions. Toward this, we build on previous work in which a stylized irrigation dilemma was simulated via a social lab experiment. Studies of the data produced from this experiment modeled participants' behavior with multiple, theoretically grounded agent-based models (ABMs). These models encode causal, human-interpretable hypotheses of decision making which generates the real-world behavior observed in the experiment. However, the accuracy of these models in fitting the experimental data is limited. Using Evolutionary Model Discovery, a recent algorithm for inverse generative social science (iGSS), we show the ability to automatically generate a wide variety of unique new ABMs which fit the experimental data more accurately and robustly than the original, manually-constructed ABMs. To do this, we algorithmically explore the space of possible behavioral rules for agents choosing how to contribute to the maintenance of the irrigation infrastructure. We find that, in contrast to the original models, our best-performing models typically have an additional element of stochasticity and favor factors such as other-regarding preferences and perceived relative income. Given that this change in just a small part of the original model has yielded such an advance, our results suggest that iGSS methods have great potential for continuing to derive more accurate models of complex social-ecological dilemmas.