Different computational strategies for different reinforcement learning problems
Pieter Verbeke, Tom Verguts, Ghent University, Belgium
Posters 3 Poster
Pacific Ballroom H-O
Sat, 27 Aug, 19:30 - 21:30 Pacific Time (UTC -7)
The Rescorla-Wagner rule remains the most popular tool to describe human behavior in reinforcement learning problems. However, several studies have shown that this learning rule is insufficient to fit human flexibility. Therefore, several extensions of the Rescorla-Wagner learning rule have been proposed. Current work investigates three of these extensions on a wide variety of reinforcement learning datasets. Specifically, we investigate the addition of (1) an adaptive learning rate, (2) modularity and (3) hierarchical learning. We observed (AIC) evidence for each additional feature if reward probabilities change in a well-identifiable manner, such as in classic reversal learning tasks. When reward probabilities are stable or when they change in a not well-identifiable manner the Rescorla-Wagner model fits the data best.