Recent emphases on differences between metaphors andsimiles pose a quandary. The two forms clearly differ instrength, but often seem to require similar interpretations. InExperiment 1 we show that ratings of comprehensibility arehighly correlated across simile and metaphor sentencesdiffering only in the presence or absence of “like”. InExperiment 2 we show that comprehensibility ratings forfigurative forms predict both early (first pass) and late(second pass) fixation durations for metaphor vehicle, butonly late fixation durations for vehicles in similes. Similevehicles appear to initially be processed similarly to literalcomparisons, with figurative interpretation occurring later.These observations are consistent with the different pragmaticstrengths, and similar interpretations of the two forms.

Successful social coordination requires being able to predict how the other people that one depends on are likely to behave. One solution to this dilemma is to establish social conventions, which constrain individuals' behavior but make prediction easier. Here, we develop a multi-agent deep reinforcement learning environment to investigate the costs associated with these conventions. In our produce-and-trade task, agents have varying production skills, but their actions must be predictable in order to be rewarded. Stronger norms improve the overall success of the group by improving the average rewards of the majority, but also systematically disadvantage agents whose specialization is in the minority of the group. Critically, this outcome is magnified by population size: as larger groups make it potentially more difficult to develop individualized representations of agents, minority agents become more likely to conform to a norm that is disadvantageous to them.

People are capable of learning diverse functional relationships from data; nevertheless, they are most accurate when learning linear relationships, and deviate further from estimating the true relationship when presented with non-linear functions. We investigate whether, when given the opportunity to learn actively, people choose samples in an efficient fashion, and whether better sampling policies improve their ability to learn linear and non-linear functions. We find that, across multiple different function families, people make informative sampling choices consistent with a simple, low-effort policy that minimizes uncertainty at extreme values without requiring adaptation to evidence. While participants were most accurate at learning linear functions, those who more closely adhered to the simple sampling strategy also made better predictions across all non-linear functions. We discuss how the use of this heuristic might reflect rational allocation of limited cognitive resources.

We test whether people flexibly shift their sampling strategy for learning a functional relationship, based on the strategy’s perceived effectiveness. While general-purpose heuristics such as gathering information evenly across the environment may often approximate optimal sampling policies when opportunities for learning are sparse, these strategies may systematically fail when much of the environment is uninformative. Across several different classes of arbitrary smooth functions, participants (N = 89) made sampling choices that were initially consistent with a simple heuristic, but shifted their sampling strategy when this heuristic failed to be informative. People were subsequently more accurate at approximating the true function for smoother functions that can be more easily predicted by sampling evenly, t(645.3) = 6.803, p < .001; nevertheless, while individuals vary considerably, aggregate judgments across all functions are very accurate, suggesting that people show inductive biases in active learning, but effectively learn when to adjust their policies.

Functional relations are prevalent in everyday life and science. Do children have intuitive knowledge of functional relations, and can they learn these relations by active information gathering (i.e., choosing a few input values and observing the corresponding outputs)? We found that 6- to 9-year-olds can learn different families of functions (linear, Gaussian, and exponential) through both informative data provided by an experimenter and data they gather from the environment for themselves. Overall, children learn linear functions more accurately than non-linear functions. When choosing data points to learn about, some children select highly similar points that only shed light on a narrow region of a function, while others choose more variable inputs and gain a more holistic view of a function. Children who use this latter, globally informative strategy have higher learning accuracy, particularly for non-linear functions. Results suggest that children are in the process of developing effective strategies for active function learning.

We investigate people’s use of strategies for sampling data in an active learning task. In the spirit of resource-rational analysis, we argue that people may often use effective heuristics to guide sampling in lieu of more computationally expensive optimization strategies, but that when they encounter evidence that their heuristics are now ineffective they flexibly shift to new strategies. When the function family changed, participants quickly updated their beliefs about the likely function family on subsequent trials. By clustering participants’ sampling behaviour, we show that people can employ varied sampling strategies, shifting strategies more often when encountering unusual function families that are more adversarial to generic sampling strategies. Not all new strategies improved participants’ performance on a subsequent prediction task; nonetheless, people’s ability to dynamically shift their active learning behaviour may help them understand the abstract features of complex relationships.

Our ability to play games like chess and Go relies on both planning several moves ahead and on recognition or gist - intuitively assessing the quality of possible game states without explicit planning. In this paper, we investigate the role of recognition in puzzle solving. We introduce a simple puzzle game to study planning and recognition in a non-adversarial context and a reinforcement learning agent which solves these puzzles relying purely on recognition. The agent relies on a neural network to capture intuitions about which game states are promising. We find that our model effectively predicts the relative difficulty of the puzzles for humans and shows similar qualitative patterns of success and initial moves to humans. Our task and model provide a basis for the study of planning and intuitive notions of fit in puzzle solving that is simple enough for use in developmental studies.

We explore whether children’s strategies on a causal learningtask show a bias observed in adults towards “exploitative” hy-pothesis revision. Adults and children (ages 4–6) were pre-sented with evidence which initially seemed to conform to asimple, salient rule (e.g. blue blocks activate a machine), butthen encountered evidence that violated this rule. The truerule in the “near” condition was more complex, but could bereached through iterative revision of the salient rule, while inthe “distant” condition, the true rule was comparatively sim-ple, but incremental revision could not yield the true rule. Par-ticipants then predicted the behaviour of a set of new blocks.Adults performed better in the near condition, while in the dis-tant condition adults did not appear to revise their initial hy-pothesis significantly. Unlike adults, children’s overall perfor-mance in both conditions was similar, while condition differ-ences may reflect a broader search for alternative solutions.