Neurocomputing: foundations of research
Exploring Adaptive Agency III: Simulating the Evolution of Habituation and Sensitization
PPSN I Proceedings of the 1st Workshop on Parallel Problem Solving from Nature
| Hi-index | 0.00 |
What simple learning rules can allow agents to cope with changing environments? We tested whether a rule that neglects base rates of events in the world -- something that is usually considered irrational -- could be as successful as Bayesian inference that combines base rates and cue accuracies - the usual standard of rationality-in making cue-based predictions about events in time-varying environments. We focused on environments in which base rates change more frequently than cue accuracies, a condition that, we argue, is common in the real world. Five strategies (Bayesian combination, cue accuracy alone, adjusted cue accuracy, base rates alone, and a Least Mean Square learning rule) were compared across "lifetimes" of 10,000 predictions, in which base rates and cue accuracy independently changed every 10, 50, 100, 500, 1000, or 5000 events. The results confirmed that simple strategies that are typically deemed irrational (base rate neglect and its opposite, conservatism) can rival the typical standard of rationality, Bayesian combination of information, by producing ecologically rational decisions in appropriately varying environments.