Decision-making with distorted memory: Escaping the trap of past experience

People generally try to learn from their past experiences. Empirically, however, we tend to draw on these experiences in a seemingly irrational way, as originally noted by psychologist Daniel Kahneman. Rather than remembering each episodic experience as a whole, people generally recall only particular aspects of those experiences, especially the most intense peaks, as well as what happened in the final moments. Kahneman referred to this as the “peak-end rule” – the recollected pain or benefit derived from an event is accurately captured by a weighted average of the most intense moment and the final parts of the episode. One important consequence of this is so-called “duration neglect,” meaning that retrospective evaluations of experiences typically do not depend on episode duration but only on the extreme and final snapshots.
As Evangelos Mitsokapas of Queen Mary University of London and LML Fellow Rosemary Harris explore in a recent paper, this peak-end rule for memory may also have an important influence over future decisions which may, for example, be more strongly influenced by the most extreme events experienced in the past, rather than by a more balanced average. In earlier work, Harris modelled such an effect using a simple discrete-choice model with decision probabilities determined by the maximum value of a random utility variable over all the past moments when the same choice was made. On the assumption that the random rewards received for each of two choices were drawn from identical distributions, this model reveals three classes of behaviour, depending on the tails of the utility distribution. Most interesting is the exponential-tail class, where there exists a particular value of noise marking a transition between a regime where the agent becomes “trapped” in one of the choices — the agent consistently samples only one of the utilities available in the long-time limit – and a regime where the agent samples both decisions. Similar examples of trapping behaviour can be found in other models explored in the behavioural economics literature.
In their recent paper, Mitsokapas and Harris extend the previous analysis to consider a generalised model in which the reward distributions for the two distinct choices are different. As they show, this has one important consequence, as the agent now faces the risk of becoming trapped in the worse choice, thereby reducing their expected long-term return. They also explore how different levels of noise affect the decision-making process and show that, for exponential distributions, there is an optimal value of noise for which the agent can always escape the trapping pitfall and choose so as to maximize its returns.
The paper is available at

Leave a Reply

Your email address will not be published. Required fields are marked *