We present a neural-network computational model of a recent experiment revealing that chimpanzees show some ability to reason probabilistically. Specifically, we show that the neural probability learner and sampler (NPLS) system can account for both success by chimpanzees and better performance by human controls. NPLS effectively combines learning probability distributions with sampling from those learned distributions to guide action choices. Because NPLS also simulates learning and use of probability distributions by human infants, this brings us closer to a unifying model of probabilistic reasoning, across various age groups and species.

Recent empirical work has provided evidence for pure reasoning in infancy, a capacity permitting flexible integration of multiple sources of information to form rational expectations about novel events (Teglas et al., 2011). However, the neural underpinnings of this capacity have remained elusive. In this work, we present the first ecologically rational, neural-level account of these findings on pure reasoning in human infants. Our work bridges two dominant approaches in computational developmental psychology, i.e. neural-network models and Bayesian modeling, substantiating the view that intuitive physics in infancy might, at least partly, involve heuristics: a set of simple, fast, resource-efficient, approximation algorithms that yield sufficiently good results.

We propose and implement a relatively simple computational neural-network model of number comparison. Training on paired comparisons of the integers 1-9 enables the model to efficiently and accurately simulate some fundamental empirical phenomena (distance and ratio effects on accuracy and response time). It also generalizes robustly to more advanced tasks involving multidigit integers, negative numbers, and decimal numbers. The work demonstrates that small neural networks can sometimes efficiently learn a powerful system that exhibits extremely robust generalization to untrained items. Some important alternate models of number comparison are considered to establish a broader context. Several predictions and suggestions are made for future empirical and computational research in this area.

Widely considered a cornerstone of human morality, trust shapes many aspects of human social interactions. In this work, we show that artificial intelligence (AI) agents learn to trust ― and rationally adapt the extent to which they trust ― by interacting with their counterparts in the trust game (Berg, Dickhaut, & McCabe, 1995), the canonical task for studying trust in behavioral and brain sciences. Leveraging reinforcement learning (RL) to train our AI agents, we systematically investigate learning trust under various parameterizations of this task and show that trust can arise from pure self interest. Our work sheds new light on the computational underpinnings of trust, taking us a step closer to developing AI agents whose values are aligned with human values.

To achieve desirable performance, current AI systems often require huge amounts of training data. This is especially problematic in domains where collecting data is both expensive and time-consuming, e.g., where AI systems require having numerous interactions with humans, collecting feedback from them. In this work, we substantiate the idea of cognitive models as simulators, which is to have AI systems interact with, and collect feedback from, cognitive models instead of humans, thereby making their training process both less costly and faster. Here, we leverage this idea in the context of moral decision-making, by having reinforcement learning (RL) agents learn about fairness through interacting with a cognitive model of the Ultimatum Game (UG; Nobandegani, Destais, & Shultz, 2020), a canonical task in behavioral and brain sciences for studying fairness. Interestingly, these RL agents learn to rationally adapt their behavior depending on the emotional state of their simulated UG responder.

Strategic decision-making is chiefly studied in behavioral economics using multi-agent games. Decades of empirical research has revealed that emotions play a crucial role in strategic decision-making, calling into question the “emotionless” homo economicus. In this work, we present a unified process-level account of a broad range of empirical findings on the effect of emotions in Prisoner’s Dilemma and Ultimatum games—the two most studied games in behavioral sciences. Under the empirically well-supported assumption that emotions modulate loss aversion, we show that Nobandegani et al.’s (2018) sample-based expected utility model can account for the effect of emotions on: (i) cooperation rate in Prisoner’s Dilemma, and (ii) the rejection rate of unfair offers in the Ultimatum game. We conclude by discussing the implications of our work for emotion research, and for developing a unified process-level account of the role of emotions in strategic decision-making.

Recent empirical work has shown that human children are adept at learning and reasoning with probabilities. Here, we model a recent experiment investigating the development of school-age children's non-symbolic probability reasoning ability using the Neural Probability Learner and Sampler (NPLS) system. We demonstrate that NPLS can accurately simulate children’s probability judgments at different ages, tasks and difficulty levels to discriminate two probabilistic choices through accurate probability learning and sampling. An extension of NPLS using a skewed heuristic distribution can also model children’s tendency to wrongly select the outcome with more favorable items but less likely to draw the favorable ones when the probabilistic choices are similar. We discuss the roles of two model parameters that can be adjusted to simulate the probability matching versus probability maximization phenomena in children, and why frequency biases children’s probabilistic judgments.

Human decision-making is filled with numerous paradoxes and violations of rationality principles. A particularly notable example is violation of cumulative independence (VoCI). Recently, there has been a surge of interest in theorizing and developing a resource-rational foundation for many such phenomena. Here we ask whether VoCI could be given a resource-rational basis too. To what extent could VoCI be explained in terms of the optimal use of limited cognitive resources? In this work, we look at VoCI through the lens of modern psychological theories of bounded rationality, presenting the first resource rational account of VoCI.We discuss the implications of our work for risky decision-making, and more broadly, human rationality.

Decades of empirical and theoretical research on human decision-making has broadly categorized it into two, separate realms: decision-making under risk and decision-making under uncertainty, with the Allais paradox and the Ellsberg paradox being a prominent example of each, respectively. In this work, we present the first unified, resource-rational account of these two paradoxes. Specifically, we show that Nobandegani et al.’s (2018) sample-based expected utility model provides a unified, process-level account of the two variants of the Allais paradox (the common-consequence effect and the common-ratio effect) and the Ellsberg paradox. Our work suggests that the broad framework of resource-rationality could permit a unified treatment of decision-making under risk and decision-making under uncertainty, thus approaching a unified account of human decision-making.

Humans not only can effortlessly imagine a wide range ofnovel instances and scenarios when prompted (e.g., a newshirt), but more remarkably, they can adequately generate ex-amples which satisfy a given set of constraints (e.g., a new,dotted, pink shirt). Recently, Nobandegani and Shultz (2017)proposed a framework which permits converting deterministic,discriminative neural nets into probabilistic generative models.In this work, we formally show that an extension of this frame-work allows for generating examples under a wide range ofconstraints. Furthermore, we show that this framework is con-sistent with developmental findings on children’s generativeabilities, and can account for a developmental shift in infants’probabilistic learning and reasoning. We discuss the impor-tance of integrating Bayesian and connectionist approaches tocomputational developmental psychology, and how our workcontributes to that research.