Abstract
Regret seems like a very negative emotion, sometimes even debilitating. However, emotions usually have a purpose, in the case of regret to help us learn from past mistakes. In this paper we first present an informal cognitive account of the way regret is built from a wide range of both primitive and more sophisticated mental abilities. The story includes Skinner-level learning, imagination, emotion, and counter-factual reasoning. When it works well this system focuses attention on aspects of past events where a small difference in behaviour would have made a big difference in outcome – precisely the most important lessons to learn. The paper then takes elements of this cognitive account and creates a computational model, which can be applied in simple learning situations. We find that even this simpli-fied model boosts simple machine learning reducing the number of required training samples by a factor of 3-10. This has theoretical implications in terms of understanding emotion and mechanisms that may cast light on related phenomena such as creativity and serendipity. It also has potential practical applications in improving machine leaning and maybe even alleviating dysfunctional regret.
Keywords: regret, cognitive model, emotion, machine learning
|
regret model architecture
Learning scaled to min/max possible scores showing better asymptotic values and faster learning
Speed up, log-scale – number of regret exposures to reach same quality of learning as simple learner