Towards unifying category learning, probability learning and risky gambling using the CAL framework of rule and attention learning
Stimulus classification is an everyday feat (e.g., in medical diagnoses by differentiating ultrasound images). Category feedback, however, is often non-deterministic (e.g., by 25% chance untrue a.k.a. probabilistic feedback) rendering experiences as somewhat unreliable, and the question is how humans (still) learn stimulus-category regularities. In probability learning and economic decisions, such as risky gambles, however, the question usually reverses to why humans do not perfectly exploit regularities when correct categorization leads to reward (e.g., non-rational probability matching; Feher da Silva, et al., 2017; Plonsky, Teodorescu, & Erev, 2015). Here, we address both questions in a domain-general framework formalizing how humans, in probabilistic tasks, learn sequential feedback regularities in parallel to visual category structures. We use our recently introduced Category Abstraction Learning (CAL) framework (Schlegelmilch, Wills, & von Helversen, 2021), a connectionist category learning model able to extrapolate and contextually modulate simple rules. We implement the idea that participants count the streak of common events (stimulus) to predict when rare events or violations of a learned rule will occur (e.g., conditional hypotheses). We show that CAL's learning mechanisms readily extend to the mentioned domains, predicting probability matching in general based, but also the proportion of strategies often discussed as Win-Stay-Lose-Shift (WSLS), and more recently studied sequential pattern learning (akin to gamblers fallacy). CAL also provides an account of expectancy priors (see Koehler & James, 2014), proposing that they stem from an awareness that unobserved stimuli lead to unobserved outcomes (contrasting) which are continuously updated during experience-based decision making. We present CAL simulations and brief reanalyses of studies on risky gambles, probability learning and fear conditioning (e.g., Szollosi et al., 2022) showing CAL's potential to address long-standing questions regarding non-stationary expectations of stimulus-outcome probabilities and risk preference in terms of rule abstraction.