The data-driven shift in modeling human reward learning

Understanding the computational basis of reward learning is critical for explaining how humans achieve goals in complex, ever-changing environments. While cognitive science has long relied on the Reinforcement Learning framework to study human learning, a recent shift towards data-driven methods is challenging this paradigm. Recent results question long-standing explanatory mechanisms, including the error-driven, incremental nature of human learning, and new methods are revolutionizing investigative approaches: data-driven methods can automate model design, construct more powerful algorithms, and search candidate mechanisms more comprehensively. But while these methods have produced highly predictive and often interpretable models that often reveal novel cognitive mechanisms, they have their own shortcomings. Future research, hence, needs to combine the strengths of both hypothesis-driven model design and data-driven discovery.