| Summary: | Constructing internal representations of the world is a fundamental aspect of cognition, allowing us to predict and control our environment. However, sensory observations are frequently noisy and incomplete, leading to the question of how such representations are acquired. We take a normative approach to this problem of inductive inference, or structure learning, asking what agents should believe about the world in light of their observations. We consider experiments from a variety of psychological domains, in each case proposing a rational structure-learning model and comparing real behaviour with model predictions. Firstly, we address the problem of rule learning in a memory-based maze task. We show that gating algorithms, a model-free approach to solving POMDPs, replicate rule acquisition in rats, as well as transfer of learning under rule reversal. Secondly, we consider the problem of behaving flexibly in environments composed of distinct behavioural regimes, or 'contexts'. Vile show that a novel decision-making model that discriminates between contexts captures a number of important animal learning phenomena including spontaneous recovery, partial reinforcement and overlearning effects, and serial reversal learning effects. Thirdly, we turn to perception and whether structure learning can explain participants' behaviour in a perceptual task. Assuming that participants aim to infer the structure underlying observed stimuli, we show that discounting of colour cues when estimating motion direction can be explained by a rational clustering model. Finally, we consider the effect of working memory capacity (WMC) on category learning. Treating category learning as a structure learning problem and modelling WMC as the quantity of inferential resources available, we replicate the positive association between WMC and both rate of learning and ability to switch between categorisation strategies. Our results suggest that the simple assumption that agents construct representations of their environments, combined with adequate modelling tools for representation and inference, can offer parsimonious explanations of behaviour in diverse areas of psychology.
|
|---|
