Introduction

The purpose of exploration is to reduce uncertainty about the aspects of one’s environment that are goal relevant or otherwise important. Yet, devising an optimal strategy to reduce uncertainty is known to be very difficult (Cohen et al., 2007; Schulz and Gershman, 2019; Sutton and Barto, 2018), especially for agents with limited memory and processing capacities. A heuristic strategy that is often efficient for exploration is focusing on the parts of the environment that one is most uncertain about. This principle of approaching uncertainty has been applied in a range of fields, including statistics (MacKay, 1992; Sebastiani and Wynn, 2000), artificial intelligence (Badia et al., 2020; Bellemare et al., 2016; Pathak et al., 2017; Raposo et al., 2021), and cognitive theories of human exploration (Schulz and Gershman, 2019; Schwartenbeck et al., 2019). Indeed, humans have been shown to approach uncertainty when learning about rewards in the environment through trial and error (Schulz and Gershman, 2019; Speekenbrink and Konstantinidis, 2015; Wilson et al., 2014; Wu et al., 2022).

However, there are also many examples of uncertainty avoidance in the decision making of humans and animals. Uncertainty avoidance has been documented in situations where resolving uncertainty may reveal negative outcomes and news (Ahmadlou et al., 2021; Botta et al., 2020; Eilam and Golani, 1989; Glickman and Sroges, 1966; Gordon et al., 2014; Gigerenzer and Garcia-Retamero, 2017; Golman et al., 2017), or may make overcoming a conflict in motivation more difficult (Carrillo and Mariotti, 2000; Golman et al., 2017). When the goal is to maximize immediate rewards, choosing the most rewarding option often entails avoiding more uncertain options (Trudel et al., 2020; Wilson et al., 2014).

How are the two conflicting tendencies to approach and avoid uncertainty reconciled when exploring? To answer this question, we must address gaps in the literature about exploration at two levels of analysis. At the computational level, it is unclear what might compel individuals to avoid uncertainty instead of approaching it, bar holding goals other than attaining knowledge. Indeed, avoiding uncertainty reduces the rate of information intake, and so might result in poorer learning. At the algorithmic level, we lack an understanding of how individuals compute uncertainty to make exploratory choices. Tallying uncertainty in an exact manner is complicated and often intractable. Several candidate algorithms for approximating the computation of uncertainty have been suggested (Schulz and Gershman, 2019), but evidence as to their use by humans is still preliminary.

It is the complexity of choosing based on uncertainty, set against the limited processing and memory capacities that are inherent to human cognition, that motivated our hypotheses regarding both the algorithmic and computational questions. First, we charted a hypothesis space of plausible algorithms for computing uncertainty and making exploratory choices (Schulz and Gershman, 2019), starting with the optimal but complex, and ending with simple approximations. Second, we hypothesized that the complexity of choosing what to explore, even when using approximate algorithms, is the key factor explaining why and when individuals might avoid uncertainty in exploration. Adhering to the goal of approaching uncertainty may well be an efficient policy for an agent with unlimited cognitive resources. Since humans have finite memory systems, inference bandwidth, and time, it stands to reason that they would try to conserve these resources by regulating their exploration (Lieder and Griffiths, 2020), possibly by selectively avoiding uncertainty. Following this insight, we examined exploratory choices as a function of two factors affecting the difficulty of making an exploratory choice: participants’ overall uncertainty about choice options (Schulz and Gershman, 2019), and forgetting.

We developed a task requiring participants to make multiple exploratory choices, incrementally building knowledge in the service of a distant goal (Figure 1). Importantly, participants were given reward feedback only at the end of a round and not after every trial, allowing us to focus on choices made to accumulate knowledge, rather than choices driven by the need to exploit available rewards. Seeking ecological validity, we designed a task that posed a challenging exploration problem for participants, requiring that they infer and remember the values of multiple latent parameters from repeated experience (Hartley, 2022; Lieder and Griffiths, 2020). The task could nonetheless be captured by a few mathematical expressions, allowing for the derivation of the optimal exploration policy. This optimal policy served as a basis for a quantitative analysis of participants’ choices and reaction times with the aim of identifying the algorithm driving their exploratory choices (Anderson, 1990; Chater and Oaksford, 1999; Waskom et al., 2019).

Results

194 participants from a pre-registered (Abir et al., 2021) sample were recruited to complete up to 22 rounds of the exploration task over four online sessions. The task simulated a room with four tables, with two decks of cards on each table (Figure 1a-b). If a card was flipped, it was revealed to be, for example, either orange or blue (each round used a different pair of colors). The proportion of orange vs. blue cards, π differed between the two decks on each table. Participants’ goal was to learn sgn(π₁ – π₂ or which deck had more orange (blue) cards on each table. We will denote this term, which serves as the learning desideratum for participants, as θ.

Figure 1.

Figure 2.

The task begins with an exploration phase, followed by a test phase. On each trial of the exploration phase participants chose which of two tables to explore, and then revealed one card from a deck on that table (Figure 1 b). Participants were instructed that the exploration phase would be followed by a test phase after a random number of trials (drawn from a geometric distribution to discourage pre-planning, Figure 1c). They were further instructed that one of the colors would be designated as rewarding at the beginning of the test phase. During the test phase, participants were asked to indicate which deck had more of the rewarding color on each table (Figure 1b). They also rated their confidence in the choice. For every correct test-phase choice they received $0.25. Crucially, they received no reward during exploration. Participants’ only incentive during the exploration phase was to maximize their confidence about the value of θ.

Three Hypothetical Strategies Derived by Rational Analysis

To explain how participants chose between tables in the exploration phase, we first asked how an optimal agent might solve the problem of choosing which table to explore on each trial of the task. We limited our consideration to strategies that optimize learning only for the next trial, since a globally optimal strategy is intractable for this task (Schulz and Gershman, 2019; Sutton and Barto, 2018). We started by deriving the optimal strategy and progressively simplified it to generate two additional strategies. While they differ in the level of complexity they assume, all three strategies direct an agent using them to approach the option they are more uncertain about.

The optimal strategy, given by the expression at the top of Figure 1d, is choosing the table affording maximal expected information gain (EIG; Gureckis and Markant (2012); MacKay (1992); Yang et al. (2016)). EIG is the difference between the uncertainty in the value of the learning desider-atum, θ, given observed cards x_{0: t} and the expected uncertainty after observing the next card on trial t +1. In other words, EIG is the amount of uncertainty resolvable on the next trial.

Figure 3.

Figure 3—figure supplement 1.

Computing the second term in the EI expression requires averaging over future unseen outcomes, which may be beyond the ability of participants. As an alternative, they might avoid computing this term by simply choosing the table they were more uncertain about at the moment of making the choice (Figure 1d, second tier; Schulz and Gershman, 2019). While this strategy has intuitive appeal, computing uncertainties may still be too complicated for human participants. An even simpler heuristic is given on the third tier of Figure 1d: choosing the table with the least prior exposure (Auer, 2002; Schulz and Gershman, 2019), measured as the number of already observed cards n_x. Since on average additional observations result in lower uncertainty, this strategy is an approximate way to approach the more uncertain table. Finally, participants might explore at random, rather than in a directed manner (Daw et al., 2006; Schulz and Gershman, 2019; Wilson et al., 2014).

Test Phase Performance Validates Observation Model

To relate the three hypothesized strategies to participants’ behavior, we assumed a model of participants’ beliefs about the goal-relevant parameter θ and the mechanism by which they updated these beliefs. We used a Bayesian observer model which forms beliefs about θ based on the actual card sequence each participant observed, and updates these beliefs according to Bayes’ rule (Figure 2). On its own, the Bayesian observer does not predict participants’ exploration choices, but only models the process of inference from observation.

Figure 4.

Figure 4—figure supplement 1.

Figure 4—figure supplement 2.

Figure 4—figure supplement 3.

Before evaluating the hypothesized exploration strategies, we sought to validate the assumptions of the Bayesian observer model. To this end, we related the predictions of the Bayesian observer model to participants’ choices during the test phase. We predicted that test accuracy should be greater when the Bayesian observer model had low uncertainty about θ at the end of the learning phase. The data supported this prediction (Figure 3). Using a multilevel logistic regression model, we confirmed that test accuracy was strongly related to the Bayesian observer’s uncertainty b=-5.59, 95% posterior interval (PI)=[-6.25,-4.95] (all effect sizes given in original units, full model and coefficients reported in Appendix 3—Table 1). Participants’ reports of confidence after making a correct choice also followed the Bayesian observer’s uncertainty b=-4.04, 95% PI=[- 4.50,-3.56]. After committing errors, participants’ reported confidence was lower overall b=-1.09, 95% PI=[-1.27,-0.92], and considerably less dependent on Bayesian observer uncertainty, interaction b=-3.10, 95% PI=[-3.76,-2.46] (Figure 3b, Appendix 3—Table 2).

Uncertainty is the Best Predictor of Exploratory Choice

To evaluate the three exploration strategies, we tested whether participants’ exploration-phase choices could be predicted from the difference between the two tables that were presented as choice options in each of the hypothesized quantities. We fit the data with a multilevel logistic regression model for each strategy (Appendix 3—Tables 3-5). In a formal comparison of the three models we found that uncertainty was the best predictor of exploratory choices, as indicated by a reliably better prediction metric (Figure 4). Accordingly, the difference in uncertainty for the table presented on the right versus the table presented on the left (Δ-uncertainty) predicts participants’ choices. Δ-EIG provides a poorer fit to choices, and Δ-exposure is anti-correlated with choice, in contradiction of the exposure hypothesis. We confirmed that our analysis approach can recover the true model generating a simulated dataset(Figure 4—figure Supplement 1). Furthermore, simulations showed that uncertainty is a sufficient predictor of choice. Simulated datasets generated by uncertainty-driven agents recreated the entire set of qualitative and quantitative results (Figure 4—figure Supplement 2). The simulations demonstrate that the surprising negative correlation between choice and Δ-exposure is an epiphenomenon of uncertainty-based exploration.

Participants Systematically Change Their Exploration Strategy According to Overall Uncertainty

We next asked whet her participants’ strategy of exploring by approaching uncertainty is modulated by the state of their knowledge when making an exploratory choice. Specifically, we examined how participants’ overall uncertainty about the two options they could choose to explore on a given trial changed the way the explored (Figure5a). Since table choice options were presented at random, participants sometimes had to choose between tables they already knew a lot about, and sometimes between tables they were very uncertain about. When overall uncertainty was high, the choice between tables had to be made with very little evidence. Note that from a normative perspective, choice should follow the difference in uncertainty between options and shouldn’t be influenced by overall uncertainty.

We found a systematic deviation in exploration strategy in relation to overall uncertainty. When overall uncertainty for the two choice options was below a certain threshold, participants chose the more uncertain table, as expected. However, when overall uncertainty was above the threshold, they chose the less uncertain table, thereby slowing the rate of information intake (Figure 5b, c).

We validated this observation using a multilevel piecewise-regression model, allowing for the influence of Δ-uncertainty on choice to differ below and above a fitted threshold of overall uncertainty. We observed a positive relationship between Δ-uncertainty and choice below the threshold b=0.97, 95% PI=[0.83, 1.11], but above the threshold we found that the influence of Δ-uncertainty on choice became strongly negative (interaction b=-4.3e+02, 95% PI=[-5.4e+02,-3.4e+02]]). The threshold was estimated to be 1.28 nats of overall uncertainty (95% PI=[1.27, 1.29]; Appendix 3— Table 6), leaving 21.58% of trials in the high overall uncertainty range (95% PI=[20.12,24.45]). This bias in exploration cannot be viewed merely as a noisier version of optimal performance. Rather, it constitutes a systematic modulation of exploration strategy on about a fifth of the trials.

Costs and Benefits of Strategically Avoiding Uncertainty

What motivates participants to systematically avoid learning about more uncertain objects? By the standards of an ideal agent, uncertainty avoidance is clearly suboptimal, as it reduces the rate of observed information, and thus the potential capacity to learn. We hypothesized that the limited processing and memory capacities that are inherent to human cognition and set it apart from the optimal agent are the reason for uncertainty avoidance. To test this hypothesis, we conduct a cost benefit analysis of uncertainty avoidance in the following sections. We ask whether uncertainty avoidance is associated with costs to learning, and whether it affords any benefits in managing cognitive effort.

Uncertainty Avoidance is not Associated with Learning Deficits

Since efficient learning is the purpose of exploration, we asked how the tendencies to approach uncertainty and avoid it when overall uncertainty is high affect learning as reflected in performance attest. If approaching uncertainty is the only rational exploration policy, then participants who tend to approach uncertainty to a greater degree should learn more and perform better at test, while participants with a strong tendency to avoid uncertainty should learn less and perform worse at test, since they are choosing to forgo valuable information as they explore.

To test these predictions we examined individual differences in exploration strategy in relation to test performance. We found that participants’ baseline tendency to approach uncertainty predicted better performance at test b=2.96, 95% PI=[2.67, 3.25] (Figure 6b; Appendix 3—Table 7). In contrast, we found no evidence that participants with a strong tendency to avoid uncertainty performed worse at test. Indeed, a stronger tendency to avoid uncertainty when overall uncertainty is high was associated with a small improvement in test performance b=1.18, 95% PI=[0.80, 1.58] (Figure 6b; Appendix 3—Table 8). Thus, modulating exploration according to overall uncertainty was not maladaptive, resulting in no decrement to learning. This result suggests that the rate of information intake is not the limiting factor for the efficiency of exploration and learning.

Figure 5.

Figure 5—figure supplement 1.

Figure 6.

Figure 6—figure supplement 1.

Strategic Exploration Involves Costly Deliberation

To understand the costs involved in exploration, we asked whether making exploratory choices in this task involves prolonged deliberation. If that is the case, and exploratory choices are guided by ΔΔ-uncertainty, we reasoned that decisions should require longer deliberation when the absolute value of Δ-uncertainty is small (Palmer et al., 2005; Shushruth et al., 2022). To test this prediction, we fit the data with a generative model of choice and RTs. We used a sequential sampling model, which explains decisions as the outcome of a process of sequential sampling that stops when the accumulation of evidence satisfies a bound. This model explains RTs as jointly influenced by participant’s efficacy in deliberating about Δ-uncertainty, and their tendency to deliberate longer vs. make quick responses (Ratcliff and McKoon, 2008; Shadlen and Kiani, 2013; Shadlen and Shohamy, 2016). One prediction of sequential sampling theory is that greater deliberation efficacy should be manifested as as greater dependence of RT on absolute Δ-uncertainty (Palmer et al., 2005).

We found that RTs indeed varied in relation to the absolute value of Δ-uncertainty as expected b=0.69, 95% PI=[0.58,0.78] (Appendix 3—Table 9). Crucially, a strong dependence of RT on the absolute value of Δ-uncertainty predicted better performance at test b=0.81, 95% PI=[0.58,1.07]. We further found that participants who tended to deliberate longer for the sake of accuracy also tended to perform better at test b=1.46, 95% PI=[0.58,2.34] (Figure 7c, Appendix 3—Table 10). In summary, participants who were better at deliberating about uncertainty during exploration, and who deliberated for longer, performed better at test. Thus, making good exploratory choices that lead to efficient learning involves prolonged deliberation.

Deliberation is Reduced by Choice Repetition

We have shown that when overall uncertainty is high participants avoid uncertainty rather than approach it, and that they do not pay a learning cost as a result. It remains to be shown that participants’ alternative strategy is able to shorten the time spent deliberating. Unfortunately, we could not test for such a benefit by directly comparing RTs as a function of overall uncertainty, as overall uncertainty is related to the difficulty of making an exploratory choice. With a single independent variable, deconfounding the effect of difficulty from the strategies used to ameliorate it is impossible. Fortunately, we could take advantage of a conceptually-related but independent tendency we observed in our dataset to examine the benefits of reduced deliberation times.

As in many learning tasks, participants in our task tended to repeat their previous choice (Wu etal.,2022),a tendency that was in dependent of Δ-uncertainty or overall uncertainty. We observed that participants generally preferred to re-choose the table they had last chosen (Figure 8b). We corroborated this with a multilevel regression model controlling for the effects of Δ-uncertainty and overall uncertainty b=0.50, 95% PI=[0.42,0.59] (Appendix 3—Table 11). Crucially, the tendency to repeat choices was also reflected in RTs, which for repeat choices were less related to Δ-uncertainty (b=-0.32, 95% PI=[-0.43,-0.22]). We also found that participants tended to make repeat choices more quickly rather than deliberate longer (b=-0.05, 95% PI=[-0.05,-0.04]; Figure 8c, Appendix 3— Table 12).

As in other aspects of exploration strategy, we observed considerable individual differences in the tendency to repeat previous choices. These differences were associated with the uncertainty based aspects of exploration discussed above (Figure 8d). Participants with a general tendency to repeat choices show stronger uncertainty avoidance when overall uncertainty is high, indicating that these two conceptually related strategies also co-occur in the population r=-0.60, 95% PI=[-0.74,-0.43] (Appendix 3—Table 11). Furthermore, the tendency to repeat previous choices is associated with better test performance, logistic regression b=0.09, 95% PI=[0.07,0.11] (Appendix 3—Table 13). The tendency to repeat is also correlated with a stronger baseline tendency to approach uncertainty r=0.32, 95% PI=[0.17,0.46] (Appendix 3—Table 11), which was shown above to be correlated with test performance. Thus, while from a normative point of view repeating the previous choice appears to be a context-insensitive heuristic, in practice participants who use this strategy do not learn any worse.

Forgetting as a Conceptual Control

Explaining participants’ deviation from the optimal exploration strategy as rational is interesting only to the extent that rationality is not a forgone conclusion. Is the alternative hypothesis of a failure in decision making also a-priori plausible?. We turned to forgetting as a second source of difficulty in our task and a conceptual control condition. Due to the random presentation of choice options, there was variability in the number of trials passed since either of the presented tables was last explored. We assumed that choosing between tables that had not been explored fora long time is more difficult than between tables for which evidence has been recently observed. Indeed, we found that RTs were longer with a larger lag, indicating greater difficulty of making a choice (log normal regression b=0.02, 95% PI=[0.02, 0.03]; Figure 9a, Appendix 3—Table 14). Furthermore, we observed that exploration choices on trials with a greater lag depended less on Δ-uncertainty b=- 0.08,95% PI=[-0.11,-0.04], and that the tendency to repeat the last chosen table on these trials was also diminished b=-0.13, 95% PI=[-0.15, -0.11] (Figure 9b, Appendix 3—Table 15). Finally, on trials with a large lag the difference in RTs between making a repeat and a switch choice disappeared, interaction b=0.02, 95% PI=[0.02,0.03] (Figure 9a, Appendix 3—Table 14). These patterns suggest that prior evidence is forgotten with increasing lag and that as a consequence exploration becomes more random. Hence, in contrast to the systematic effect of overall uncertainty, forgetting results in a failure to make principled exploratory choices.

Figure 7.

Figure 7—figure supplement 1.

Figure 8.

Figure 8—figure supplement 1.

Figure 9.

Figure 9—figure supplement 1.

Discussion

We examined the cognitive computations behind exploratory choices using a paradigm that encourages incremental learning in the service of a distant goal. We found that uncertainty played an important role in guiding participants’ choices about how to sample their environment for learning. In general, participants chose to learn more about the options they were more uncertain about. However, when overall uncertainty was especially high, participants instead avoided the more uncertain options and sampled the options they already knew more about. In addition, we found that participants tended to repeat previous choices. Together, this pattern suggests that participants systematically balance approaching and avoiding uncertainty while exploring.

Examining individual differences in exploration and learning revealed the costs and benefits of avoiding uncertainty when exploring. We found that strategically avoiding uncertainty is not associated with a detriment to learning, even though it slows down the rate of information intake. We also found an association between the length of deliberation and learning efficiency. Participants who deliberated longer also learned better, and deliberation time could be shortened by repeating previous choices. Based on these results, we conclude that balancing approaching and avoiding uncertainty is a way to manage cognitive resources by regulating deliberation costs. In this sense, our results serve as an example of how human cognition is adapted to the inherent constraints of the human mind, consistent with the resource rationality framework (Lieder and Griffiths, 2020).

While the literature on exploration is expansive, the paradigm presented here extends it in important ways. Researchers of reinforcement learning have previously examined how exploration manifests when agents learn incrementally about their environment. Crucially, this literature has focused on cases where reward can be gained on each trial (Brown et al., 2022; Cohen et al., 2007; Daw et al., 2006; Schulz and Gershman, 2019; Song et al., 2019; Tversky and Edwards, 1966; Wilson et al., 2014; Wu et al., 2022). In contrast, our task was designed to remove the impetus to exploit current knowledge immediately, a motivation that predominates exploration in tasks with immediate reward. Accordingly, we were able to observe many exploratory choices and had greater experimental power to describe in detail how participants approach uncertainty and when they avoid it instead. Secondly, exploration has been studied in the information search literature (Gureckis and Markant, 2012; Markant and Gureckis, 2014; Oaksford and Chater, 1994; Petitet et al., 2021; Rothe et al., 2018; Ruggeri et al., 2017). In most studies of this field participants make decisions without relying on their memory, as the entire history of learning is displayed to them on screen (cf. related work in active sensing; Yang et al., 2016). This differs from our task, which places heavy demands on memory. Rather than treating capacity limitations as a source of noise and a nuisance to measurement, we find that the rational use of limited resources is central for successful exploration.

Several previous studies of exploration inspired us to design a task with separate exploration and test phases. Using the “observe or bet” paradigm, Tversky and Edwards (1966) examined how participants trade off exploration and exploitation on a trial-by-trial basis. By using a short block of exploration followed by a test, Wilson et al. (2014) achieved a first reliable demonstration of directed exploration in humans. Finally, the expansive literature on the description-experience gap (Wulff et al., 2018) has used a similar paradigm to examine when participants choose to self terminate their exploration, and how that affects their learning. The paradigm presented here extends these approaches, as it is crafted to reveal the strategy driving each exploration choice.

We observed considerable individual differences in exploration strategy, as would be expected in a complex task requiring memory-based learning and inference. In the face of such variability, one may question the prudence of drawing conclusions about the population, since the average might be a poor summary of a plurality of idiosyncratic strategies. However, the strong correlation we observed between individual differences in exploration and test performance mitigates this concern. The correlation suggests that participants who were engaged with this task and able to learn from observation can well be described as exploring by a strategy of combining approaching and avoiding uncertainty. The relationship between test performance and RTs lends additional mechanistic support to this idea.

Our theoretical analysis and experiments leave several questions open. First, overall uncertainty in our task was correlated with the number of cards observed. While our results hold when trial number is added as a covariate to the regression models (see Appendix 3—Table 16), future work orthogonalizing overall uncertainty and time on task would help to fully disentangle the contribution of each factor to uncertainty avoidance.

Another open question is the nature of the limitation driving participants to avoid uncertainty when overall uncertainty is high. This could be due to limitations in committing prior experiences to memory, inferring latent parameters from disparate experiences, retrieving prior knowledge, or estimating the uncertainty of existent knowledge. While the idea that decisions based on high overall uncertainty are more difficult has been raised previously (Schulz and Gershman, 2019; Shafir, 1994), an explanation grounded in cognitive mechanisms is still needed. Accordingly, the mechanism by which uncertainty avoidance ameliorates choice difficulty remains unknown.

One intriguing explanation for the source of difficulty and the way it is managed lies in the distinction between strategies dependent on remembering single experiences and those dependent on the incremental acumulation of knowledge in the form of summary statistics (Collins et al., 2017; Collins and Frank, 2012; Daw et al., 2005; Knowlton et al., 1996; Plonsky et al., 2015; Poldrack et al., 2001). Both strategies could contribute to performance in tasks such as ours. A participant may be encoding prior observations as single instances, or summarizing them into a central tendency with a margin of uncertainty around it. Crucially, each strategy is associated with a different profile of cognitive resource use. Keeping track of individual experiences is much costlier than tracking a single expectation and a confidence interval around it (Daw et al., 2005; Nicholas et al., 2022) and more likely to incur costs when switching between exploring different tables. Prior work suggests individuals use single experiences or summary statistics according to the reliability of each strategy, and the cost of using it (Daw et al., 2005; Nicholas et al., 2022). In our case, summary statistics may be perceived as unreliable when overall uncertainty is high, compelling participants to rely on committing individual experiences to working memory (Bavard et al., 2021; Collins et al., 2017; Daw et al., 2005; Duncan et al., 2019; Poldrack et al., 2001). Furthermore, recent work examining how humans make a series of dependent decisions demonstrates that the tension between remembering single experiences and discarding them in favor of summary statistics is accompanied by a tendency to revisit previous choices instead of switching to new alternatives (Zylberberg, 2021).

The questions we addressed here were partly motivated by the well-established observation that humans and animals often avoid uncertainty in various situations. Two broad categories of explanation for such avoidance have been proposed (Golman et al., 2017). First, individuals avoid resolving uncertainty when it could lead to negative news, for example by avoiding ambiguous prospects when making economic choices (Ellsberg, 1961; Fox and Tversky, 1995). An extension of this idea is dread avoidance (Gigerenzer and Garcia-Retamero, 2017; Golman et al., 2017). One might avoid resolving the uncertainty about a medical diagnosis to avoid the unpleasant affective response to the news, even if the information could be very useful in determining treatment. Relatedly, humans might avoid uncertainty as a by-product of pursuing a goal other than exploration. More uncertain options are often avoided for the sake of choosing immediately rewarding options (Trudel et al., 2020; Wilson et al., 2014) — why try an unknown dish, when your absolute favorite is on the menu? Lastly, uncertainty avoidance may be a strategy for managing conflict between different motivations, or different mechanisms of action selection (Carrillo and Mariotti, 2000; Golman et al., 2017). For example, to maintain their diet, an individual might choose to avoid resolving the uncertainty about what snacks can be found in the office kitchen. Our findings highlight a different kind of strategic uncertainty avoidance. In our tasks there were no negative consequences to learning about the color proportions of card decks, and no conflicting motivations. Rather, we explain participants’ tendency to avoid uncertainty in terms of managing their limited cognitive resources.

The idea of a balance between approaching and avoiding uncertainty has conceptual parallels in other literatures. A group of relevant findings concern how animals explore their proximal environment. A classic finding in rats is that when placed in a novel open arena, they alternate between the exploration strategy of walking around the arena (uncertainty approaching) and a strategy of returning to their initial position and pausing there (termed “home base” behavior, which is uncertainty avoiding; Eilam and Golani, 1989). Relatedly, by using computational models to understand how rats use their whiskers to explore near objects, researchers have identified an alternation between uncertainty approaching and avoiding strategies (Gordon et al., 2014). Recent work in mice and primates has uncovered neural circuits driving exploration by framing the problem of exploration as striking a balance between approach and avoidance (Ahmadlou et al., 2021; Botta et al., 2020; Ogasawara et al., 2022). Our findings highlight the shared computational principles between human exploration in symbolic space and animal exploration of the physical environment and suggest that mechanisms involved in avoidance responses may also play a part in knowledge acquisition.

Finally, planning (Hunt et al., 2021), learning (Gureckis and Markant, 2012), and sensing (Gordon et al., 2014; Yang et al., 2016) are increasingly studied as active processes, situated within our environment and interacting with it. Understanding the complicated dynamics between agent and environment has been greatly facilitated by comparing behavior against the computational ideal of maximizing the amount of information observed (Oaksford and Chater, 1994; shman, 2019; Schwartenbeck et al., 2019). The findings we present here suggest a modification to this computational premise. Rather than trying to uncover as much information as possible, the goal of human exploration may be to maximize the amount of information retained in memory, by modulating the rate and order of observed information.

Methods

Data Collection and Participants

A sample of 298 participants was recruited via Amazon MTurk to participate in four sessions of the exploration task. They were paid $3.60 for each session and earned a bonus contingent on their test phase performance, adding up to $4.50 for the first session and $6 for later sessions. Additionally, a $2 bonus was paid out for completion of the fourth session. Participants were asked to complete the four sessions over the course of a week and were invited by email to each session after the first, as long as the data from their last session was not excluded according to the criteria we had specified (see below). All participants provided informed consent; all protocols were approved by the Columbia University Institutional Review Board.

The first session was terminated early for 89 participants due to recorded interactions with other applications during the experiment or failure to comply with instructions. An additional 32 sessions played by participants who had successfully completed the first session were excluded for the same reasons. One participant was excluded after reporting technical problems with stimulus presentation in the second session. Twenty-seven further sessions were excluded for failure to sample cards from both decks, a prerequisite for learning on which participants were instructed as part of the training. Altogether, data from 194 participants was included in the analyzed sample (120 female, 72 male, 2 other gender, average age 29.63, range 20-48). This sample included 194 first sessions, 156 second sessions, 129 third sessions, and 116 fourth sessions.

Before running this experiment, we pre-registered (Abir et al., 2021) a sample size of 190 participants satisfying our exclusion criteria. We chose this number to be three times larger than a preliminary sample of N=62 participants, which provided the dataset we used to develop our analysis approach and pipeline, and first identify exploration strategies as described above. Results for the preliminary sample are provided in figure supplements.

Task Design and Procedure

On each round of the exploration task participants were presented with a simple environment of four tables with two decks of cards on each table. Tables were distinguished by unique colorful patterns and decks by geometric symbols that did not repeat within an experimental session. The hidden side of each card was painted in one of two colors, with a unique color pair for each round. The proportion of colors in each deck were determined pseudo-randomly (see Supplementary Information), resulting in variability in the difference in proportion between each deck pair - the learning desideratum of this task.

At the beginning of each round, participants were first presented with the color pair for the round, and then with the table-deck assignments. Participants then had to pass a multiple-choice test on the table-deck assignment, making sure they remembered the structure of the task before proceeding to explore. Failing to get a perfect score on this test resulted in repeating this phase. The exploration phase then commenced. Trial structure for the exploration phase is depicted in Figure 1b. The lengths of the exploration phases varied from round to round. They were sampled from a geometric distribution with rate, shifted by 10 trials. The same list of round lengths was used for all participants, but their order was randomized.

Following the exploration phase, participants were tested on their learning. They were presented with the rewarding color for this round, and then had to indicate which deck had a greater proportion of that color on each table (Figure 1c). After answering this question for each of the four tables, they rated their confidence in each of the four choices on a 1-5 Likert scale. Participants were then told whether each of the test choices were correct, and the true color proportions for the two decks on each table were presented to them as 10 open cards.

The first session started with extensive instructions explaining the structure of each of the two phases of the task and clearly stating the learning goal. Participants were also instructed on the independence of color proportion within each deck pair, necessitating sampling from both decks to succeed in the task. The instructions also included training on how to make the relevant choices in each of the two stages. A quiz followed the instruction phase, and participants had to repeat reading the instructions if they had given the wrong response to any question on this quiz.

Each session started with a short practice round (12-19 trials). Data from this round was excluded from analysis. In the first session participants then played three more rounds and in later sessions five more rounds, for a total of 18 experimental rounds.

Data Analysis

Analysis was performed using Julia 1.4.2. Hierarchical regression models were fitted using the Stan probabilistic programming language 2.30.1 (Carpenter et al., 2017), using the interface supplied by the brims package version 2.16.1 (Burkner, 2017), running on top of R 4.1.2. The complete computing environment was packaged as a Docker image, which can be used to reproduce the entire analysis pipeline. Sequential sampling models were fitted on a separate Docker image (Chuan- Peng et al., 2022) containing HDDM 0.8 (Wiecki et al., 2013) running on top of python 3.8.8.

Bayesian Observer

Each of the three hypothesized strategies for exploration postulates a different summary statistic of prior learning as the driver of exploratory choice. To derive these summary statistics, we first had to construct a model of prior learning. We chose a simple Bayesian observer model (Behrens et al., 2007; Yang et al., 2016). Like our participants, this model’s goal was to learn θ = sgn(π₁ – π₂) from observed outcomes x_0:t. It did so by placing a probabilistic prior over the value of each updating it after every observation according to Bayes’ rule, and solving for θ using the rules of probability. The result is a posterior distribution capturing the agent’s expectation of the value of θ, and their uncertainty about the expectation. This process is depicted in Figure 2 for two tables and their matching pairs of decks.

This computation can be put into formulaic form as follows. At the beginning of a round, the Bayesian observer places a flat Beta distribution prior on the proportion of colors in each of the eight decks:

After observing a card, this prior would be updated to form a posterior distribution. Since the posterior of a Beta prior and a Bernoulli observation likelihood is also a Beta distribution, the posterior has a simple analytic form: after completing t trials, observing c_i cards of one color and t – c_i cards of the other color, the posterior would

We can then find the probability that θ = 1, i.e. that π₁ > π₂, by calculative the probability that π₂ is smaller than a given π₁ = z, and integrating over z, the possible values of π₁:

Where f is the Beta probability density function, F is the Beta cumulative density function, and x_0:t are observations thus far. We computed the value of this integral numerically using the Julia package QuadGK.jl. Finally, θ can only take two values, and so

Computing Hypothesized Decision Variables

The theory of decision making defines a decision variable as the quantity evaluated by the decision maker in order to choose between two choice options (Shadlen and Kiani, 2013). The difficulty of the decision should scale with the absolute value of the decision variable. Each of the three hypothesized strategies is defined by a specific summary statistic of prior learning that might serve as the decision variable for an exploratory choice. The three summary statistics are given in Figure 1e.

Both EIG and uncertainty are derived from the uncertainty of the posterior for θ as defined above. We quantified uncertainty as the entropy of the posterior belief (MacKay, 1992; Oaksford and Chater, 1994; Yang et al., 2016).

Entropy takes the unit of nats, ranging from 0 should the participant be absolutely sure about the value of θ for both table choice options, to 0.69 when they know nothing about a table. This is the equivalentof1bitofinformation, were we to replace the natural logarithm with a base 2logarithm.

Estimating Multilevel Bayesian Models for Inference

The regression coefficients and PIs reported here were all estimated using multilevel regression models accounting for individual differences in behavior. We used regularizing priors for all coefficients of interest to facilitate robust estimation (Table 1). For RT data we selected informative priors for the interceptterm in the regression (capturing the grand average of RTs)following established recommendations (Schad et al., 2021). For predicting choices, we used logistic regression, for confidence ratings we used ordinal-logistic regression, and for average RTs we used log normal regression. We estimated these models with Hamiltonian Monte Carlo implemented in the Stan probabilistic programming language using the R package brms. Three Monte-Carlo chains were run for each model, collecting 1000 samples each after a warm up period of at least 1000 samples (warm up was extended if convergence had not been reached). Sequential sampling models were estimated using slice sampling, implemented in the python package HDDM. Four Monte-Carlo chains were run for each model, collecting 2000 samples each after a warm up period of at least 2000 samples. Convergence for both model types was assessed using the Ȓ metric, and visual inspection of trace plots. R syntax formulae and coefficients for covariates for all models mentioned in the main text are reported in Supplementary Information.

Sequential Sampling Model of Reaction Times

To draw inference from participants’ RTs we turned to the sequential sampling theory of deliberation and choice. This theory encompasses a family of models in which decisions arise through a process of sequential sampling that stops when the accumulation of evidence satisfies a threshold or bound (Palmer et al., 2005; Shadlen and Kiani, 2013). From this family of models we chose to use the drift diffusion model (DDM) to fit our data, as it is very well described and extensively studied (Ratcliff and McKoon, 2008; Shadlen and Kiani, 2013). The DDM explains RTs as the culmination of three interpretable terms. The first is the efficacy of a participant’s thought process in furnishing relevant evidence for the decision - in our case the efficacy of calculating Δ-uncertainty (the drift rate in DDM parlance). The second term governs the participant’s speed-accuracy tradeoff by determining how much evidence they require to commit to a decision. This can also be thought of as how long a participant is willing to deliberate when a decision is difficult(bound height). Finally, the portion of the RT not linked to the deliberation process is captured by a third term (non-decision time). Since behavior was considerably different when overall uncertainty was high, DDM models were fit excluding trials with total uncertainty above the participant’s estimated threshold.

Table 1.

Model Evaluation

We compared the models of choice and RTs to alternative models, either reduced or expanded (see Supplementary Information). We used the LOO R package to perform approximate leave- one-out cross validation for models implemented in Stan. This method uses pareto-smoothed importance sampling to approximate cross validation in an efficient manner (Vehtari et al., 2017). Models implemented in HDDM were compared using the DIC metric. We also performed recovery analysis and posterior predictive checks for our models, making sure they capture the theoretically important qualitative features of the data.

Acknowledgements

We thank the Shohamy lab, Christopher A. Baldassano, and Gabriel M. Stine for their insightful discussion of the project. We are thankful for the support of the Stan user community, especially Matti Vuorre and Paul Burkner. We are grateful for funding support from the NSF (award #1822619 to D.S.), NIMH/NIH (#MH121093to D.S.)and the Templeton Foundation (#60844 to D.S.).

Additional Information

Author Contributions

Y.A., M.N.S, and D.S. designed research; Y.A. collected and analyzed data; M.N.S and D.S supervised.

Appendix 1

Full Details of Task and Procedure

Recruitment

Participants were recruited from the pool of Amazon Mechanical Turk (MTurk) vetted by cloudresearch.com (Hauser et al., 2022; Litman et al., 2017). We further restricted enrollment to participants with an approval rating higher than 95%, and at least 100 prior jobs completed. Enrollment was also restricted to participants registered as being 18-35 years old. Participants were presented with an ad for a multi-session study, describing base pay and the performance-dependent bonus. The ad stated that we were only looking for participants who were willing to complete all four sessions, and that the task required undivided attention.

After accepting the task, participants were directed to a website running the experiment, which was coded using jsPsych 6.0.4 (De Leeuw, 2015).

Instructions and Training

At the beginning of the first session, participants were thoroughly instructed about the task, and practiced each of its phases. First, the two-phase structure and the learning goal were introduced. Participants then were shown an example table with two decks on it. They practiced choosing a deck on a single table. Participants were instructed that only the symbol on the deck determined its identity, while its location (which was randomly determined) did not matter. Participants then completed ten practice trials limited to choosing decks on a single table, with the goal of figuring out the difference in proportions of colors between the decks. In this practice, one deck had 7/10 cards of the rewarding color, and the other 9/10 cards. After practicing choosing decks, participants were told which had more of the rewarding color, and it was explained that the differences could be more difficult to figure out in the game itself. Next, participants were told that while both decks may have a majority of cards of the same color. As such, they would have to sample from both decks to learn which had more of each color. This point was demonstrated by presenting ten cards from two decks -and asking (i) which deck had a majority of color 1 cards (both did), (ii) which had a majority of color 2 cards (none did), (iii) which deck had more of color 1 than the other (one of the decks did), and (iv) which deck had more of color 2 than the other (the other deck did). A failure to give a correct answer on all four questions prompted a repetition of this section.

Participants were then introduced to choosing a table before making a deck choice. As practice, they were tasked with revealing a single card from each deck on two tables over 4 trials. Failing to do so resulted in participants having to repeat this section.

Next, the test phase was introduced. Participants were instructed that a particular deck had more of the rewarding color in it. They then had to choose that deck on the following test screen. Having successfully done so, they received the same visual feedback they would receive in the actual task - ten cards of each deck were presented to them, demonstrating the true proportion of colors in each deck. A message was displayed alongside this demonstration, stating the accuracy of their choice.

Before starting work on the main task, participants were reminded that the test phase could commence on any given trial. They then answered six multiple-choice questions about the structure of the task and their goal. If they got any of these wrong, they were instructed on the correct answer, and then had to repeat the quiz. After successfully completing the quiz, they played the whole task over a 20-trials-long practice round. For this practice, only two tables were included in the set of stimuli. Finally, after completing this practice round, they continued to play four more rounds of the full task with four tables.

Each of the following three sessions began with a short reminder of the instructions. Participants then completed five rounds of the task. The first round was always a short one (12-19 trials, these are the four lowest numbers drawn from the geometric distribution of round lengths), and was treated as a practice round in analysis, i.e. data from this round was discarded.

Procedure for Single Round

Familiarization

At the beginning of each round, participants were presented with each of the tables included in the round, and the two decks associated with each table. They were then tested on these associations: they were shown a table, and two decks next to it, and had to indicate which of the two belonged to the table. Failure to answer correctly on all eight trials resulted in a repetition of the introduction. Following the introductions of tables and decks, participants were shown the two card colors used in this round. They were reminded about their goal, and then commenced to play the learning phase.

Learning Phase

Each trial of the learning phase began with a 500ms period during which a fixation cross was displayed at the centre of the screen. Two tables were then presented as choice options. Choice options were chosen for each trial at random, and the left-right presentation of the two choice options was also determined randomly. Participants indicated their choice of table by pressing either the ‘d’ or ‘k’ keys. Next, the unchosen table was removed from the screen, and a frame with the same pattern as the chosen table was presented for a period of 1000 ms. This was followed by a 500ms presentation of a fixation cross at the center of this frame. Then, the two decks associated with the table appeared in the frame (deck location was determined randomly). Using the same keys, participants chose to reveal a card on one of the decks. A short animation was played showing the deck being shuffled, and a single card was flipped to reveal its color. The colored card remained on screen for 2400 ms. A 1700 ms inter-trial-interval followed each trial.

Test Phase

At the end of the learning phase, participants were instructed that they will now be tested on their learning, and were shown the color designated as the rewarding color for this round. They then chose the deck they believed had more of the rewarding color on each table. After making all four choices, participants were shown their choice on each table and asked to rate their confidence that their choice was indeed the correct one.

Memory Test

Following the confidence ratings, participants were tested on their memory of table and deck associations. They were shown each deck participating in the round, and had to indicate which of the four tables it belonged to. Participants received no feedback for this memory test.

Feedback

Next, participants received feedback for their test-phase choices. For each table they were shown ten cards drawn from each deck. The cards represented the true proportion of colors in the deck. They were reminded which deck they had chosen, and were told whether that was the correct choice. After observing this for each of the four tables, participants were told how much bonus money they earned in this round.

Debriefing

At the end of the experiment, after collecting demographic information, participants were asked about any strategy they may have implemented to remember the card colors better, to choose between decks and between tables in the learning phase, and between decks in the test phase. Lastly they were asked if anything in the instructions remained unclear. These responses were evaluated for any use of external aids, such as pen and paper, and for any technical difficulties.

Compliance and Attention Checks

We implemented several measures to incentivize participants to devote their undivided attention to the task. First, the task ran in full screen mode, and if participants chose to exit it, a warning message was shown explaining that this study only runs in fullscreen mode. Additionally, refreshing the webpage and right-clicking on it were disabled. Instances where the participant interacted with another application on their computer were recorded by jsPsych, and when this occurred a warning message was displayed on screen.

Participants had to make each learning-phase choice within 3000 ms. Failing to do so resulted in the display of a warning message asking them to choose more quickly. Additionally, if a reaction time (RT) of less than 250 ms was recorded on three consecutive choices, a warning message asking participants to comply with instructions was displayed on screen.

When more than ten warning messages of any kind had been displayed, the session terminated, and participants were asked to return the job to MTurk. Participants were paid for terminated sessions, but were not invited to following sessions.

Appendix 2

Preliminary Sample

Data Collection and Participants

A preliminary sample of 70 participants was recruited via MTurk to participate in four sessions of the task. Task design was identical to that later used for the pre-registered sample. Recruitment was not limited to the cloudresearch.com approved sample, since data collection predated the proliferation of fake worker accounts on MTurk (Hauser et al.,2022).

The first session was terminated early for 7 participants due to recorded interactions with other applications during the experiment, or failure to comply with instructions. An additional 8 sessions played by participants who had successfully completed the first session were excluded for the same reasons. Five further sessions were excluded for failure to sample cards from both decks, a prerequisite for learning on which participants were instructed as part of the training. Altogether, data from 62 participants was included in the analyzed sample (33 female, 28 male, 1 other gender, average age 29.10, range 20-38). This sample included 62 first sessions, 45 second sessions, 33 third sessions, and 29 fourth sessions.

Results

Results from the pre-registered sample largely replicate the results we first observed in the preliminary sample (see matching figure supplement for each figure). Two points of divergence are observed. In the preliminary sample, we didn’t see a significant correlation between the tendency to avoid uncertainty when overall uncertainty is high and testperformance, while in the larger pre-registered sample we observed a positive significant correlation. This difference could be due to sample size, as the correlation is weak, we shouldn’t over interpret it.

A second point of divergence regards the plotting of predicted RTs by test performance tertile. While both in the preliminary sample and the pre-registered sample we find that the bound height estimated by the DDM is predictive of test performance(Supplementary Table 12), in the preliminary sample this relationship does not translate to a monotonic rising of bound height when comparing participants by test performance tertile (Supplementary Figure 4b). We hesitate to interpret any non-linear relationship between the bound height and test performance, given the small size of the sample, the divergence from the pre-registered sample, and the complexity of the models involved.

Appendix 3

Table 1.

Table 2.

Table 3.

Table 4.

Table 5.

Table 6.

Table 7.

Table 8.

Table 9.

Table 10.

Table 11.

Table 12.

Table 13.

Table 14.

Table 15.

Table 16.

Table 17.