NIMITEK Consistent temporal order speeds association learning: reinforcement learning
Why are unrelated associations learned more rapidly in a consistent temporal order? Observers viewed highly distinguishable, fractal objects and learned by trial and error to respond to each object with a particular motor response (one of four). In five experiments, associations between visual objects and motor responses were learned more rapidly for objects presented in a consistent temporal order (i.e., objects with consistent predecessor objects). Incremental learning of action weights for current and past objects does not account for the observed effects of temporal order ( direct actor ). However, a modified model with differential learning rates for current and past objects agrees qualitatively with observations. In the modified reinforcement model, a Kalman filter quantifies the certainty with which past observations predict future rewards and adjusts learning rates accordingly (Sutton, 1992, Proceedings of the 7^th Yale Workshop on Adaptive and Learning Systems, pp. 161-166). But does reinforcement learning of additional action weights truly capture the essence of the temporal order effects? We also consider an alternative view, according to which consistent temporal order eases the recognition problem posed by unfamiliar fractal objects.
NIMITEK, Reinforcement Learning
Die Daten werden geladen ...
Keine Ergebnisse gefunden, bitte ändern Sie Ihre Suchanfrage.