Range-adapting reinforcement learning and memory

A wealth of evidence in perceptual and economic decision-making research suggests that the subjective value of one option is determined by the range of the values of other available options (i.e., the context).

Yet, up until today, this literature has almost exclusively focused on decision situations where contexts are explicit and static, thus ignoring more ecological and realistic scenarios where value ranges are dynamically shaped by past outcomes.

This proposal will investigate if and how range adaptation also affects the computation of outcome values that are acquired from experience and stored in memory.

We hypothesize that range-adaptation is pervasive in learning, because it addresses a fundamental computational cost-benefit trade-off.Indeed, we propose that, although range adaptation is locally optimal, because it optimizes information processing by tuning the value response function to the experienced outcome distribution, this comes at the price of forfeiting the capacity to generalize learned-values to new contexts.

The proposed research will design and validate a comprehensive behavioural and computational framework to test this hypothesis.

More precisely, combining original experimental paradigms, behavioural analyses and computational modelling, we will i) provide a clear algorithmic description of outcome range adaptation, ii) elucidate the nature of the distorted memory representations of outcome values, and, finally, iii) probe the pervasiveness and adaptive value of this process empirically (by running experiments in different cultures) and theoretically (through extensive simulations).

We believe this project has the potential to further elucidate new fundamental computational constraints of human learning and decision-making.

Our theoretical framework and empirical results will undoubtedly contribute to the debate about the ultimate sources of human bounded rationality, with potential impact beyond cognitive science.