Insight and strategy 1 Running head: INSIGHT AND STRATEGY IN MULTICUE LEARNING Insight and Strategy in Multiple Cue Learning

In multiple-cue learning (also known as probabilistic category learning) people acquire information about cue-outcome relations and combine these into predictions or judgments. Previous studies claim that people can achieve high levels of performance without explicit knowledge of the task structure or insight into their own judgment policies. It has also been argued that people use a variety of suboptimal strategies to solve such tasks. In three experiments we re-examined these conclusions by introducing novel measures of task knowledge and self-insight, and using ‘rolling regression’ methods to analyze individual learning. Participants successfully learned a four-cue probabilistic environment and showed accurate knowledge of both the task structure and their own judgment processes. Learning analyses suggested that the apparent use of suboptimal strategies emerges from the incremental tracking of statistical contingencies in the environment. Insight and strategy 3 Insight and Strategy in Multiple Cue Learning A fundamental goal of cognition is to predict the future on the basis of past experience. In an uncertain world this involves learning about the probabilistic relations that hold between the available information and an outcome of interest, and integrating this into a singular judgment. Thus a stock-broker draws on various market indicators to predict the future price of a share, and a race-course pundit uses factors such as form and track condition to pick the likely winner of a race. The underlying structure of this kind of prediction is captured in the multiple cue learning framework (Brunswik, 1943; Hammond, 1955), which focuses on how people learn from repeated exposure to probabilistic information. This paradigm, also known as probabilistic category learning, has been applied in numerous areas of psychology, including human judgment (Brehmer, 1979; Doherty & Kurz, 1996; Klayman, 1988; for a review see Goldstein, 2004), learning and memory (Gluck & Bower, 1988; Knowlton, Squire & Gluck, 1994; Shanks, 1990), neuroscience (Ashby & Ell, 2001; Poldrack et al., 2001) and social cognition (Gavanski & Hoffman, 1986; Rappoport & Summers, 1973). A prototypical experimental task is the weather prediction task (Knowlton et al., 1994). In this task people learn to predict a binary outcome (rainy or fine weather) on the basis of four binary cues (four distinct tarot cards, see Figure 1). Each card is associated with the outcome with a different probability, and these combine to determine the probability of the outcome for any particular pattern of cards. The trials in a task are made up of a representative sampling of possible card combinations. On each trial participants see a specific pattern of cards, predict the weather, and receive feedback as to the correct outcome. This enables them to gradually learn the cue-outcome relations, and thus improve the accuracy of their predictions. Insight and strategy 4 There has been extensive research on the conditions under which people are able to master such tasks (Brehmer, 1980; Doherty & Balzer, 1988; Klayman, 1988). The main findings are that people perform well when the cues are few in number and linearly related to the outcome, when the content of the problem is meaningful, and when they receive sufficient trials and appropriate feedback. An important question that has received less attention concerns the relation between people’s performance and their knowledge of what they are doing. The Apparent Dissociation of Learning and Insight An intriguing finding emerging from recent studies is that even when people perform well in such tasks, they seem to lack insight into how they achieve this (Evans, Clibbens, Cattani, Harris, & Dennis, 2003; Gluck, Shohamy & Myers, 2002; Harries, Evans & Dennis, 2000; Wigton, 1996; York, Doherty & Kamouri, 1987). This is illustrated in Gluck et al.’s (2002) study with the weather prediction task (hereafter WP task). They found that although participants attained high levels of predictive accuracy they demonstrated little explicit knowledge about what they were doing. In particular, in questionnaires administered at the end of the task they gave inaccurate estimates of the cue-outcome probabilities, and there was little correspondence between self-reports about how they were learning the task and their actual task performance. This apparent dissociation between learning and insight has been taken as evidence for two separate learning systems (Ashby et al., 2003; Knowlton et al., 1996; Reber & Squire, 1999; Squire, 1994): (i) an implicit (or procedural) system that operates in the absence of awareness or conscious control, and is inaccessible to self-report; (ii) an explicit (or declarative) system that requires awareness and involves analytic processing. Tasks like the WP task, which require the gradual learning and integration of probabilistic information, are Insight and strategy 5 generally considered to involve (i). The lack of self-insight on this task is thus explained by the operation of an implicit system to which participants lack access. It is also argued that these two systems are subserved by distinct brain regions that can be differentially impaired (Ashby et al., 2003; Knowlton et al., 1996; Poldrack et al., 2001). Thus the WP task has been used to reveal a distinctive pattern of dissociations amongst patient populations. For example, Parkinson’s disease patients with damage to the basal ganglia show impaired learning on the task, despite maintaining good explicit memory about task features (Knowlton, Mangels & Squire, 1996). In contrast, amnesic patients with damage to the medial temporal lobes appear to show normal learning but poor declarative memory of the task (Knowlton et al., 1996; Reber, Knowlton & Squire, 1996). If correct, these conclusions have wide repercussions for everyday reasoning, and for the understanding and treatment of patients with neurological damage. However, there are several reasons to be cautious about this dual-process framework. In this paper we will focus on two fundamental issues: the measurement of insight and the analysis of individual learning. The Measurement of Insight It is important to distinguish between a learner’s insight into the structure of a task (task knowledge) and their insight into their own judgmental processes (self-insight). In the case of the WP task, this translates into the difference between a learner’s knowledge of the objective cue-outcome associations, and their knowledge of how they are using the cues to predict the outcome. And there is no guarantee that the two coincide. Someone might have an incorrect model of the task structure, but an accurate model of their own judgment process. Politicians seem particularly prone to this. Previous research tends to run the two together. Thus it is not always clear whether claims about the dissociation between insight and learning refer to a dissociation between Insight and strategy 6 self-insight and learning, task knowledge and learning, or both. Further, this conflation can infect the explicit tests given to participants. Questions that are designed to tap someone’s insight into their own judgment processes may instead be answered in terms of their knowledge about the task. Such confusion needs to be avoided if firm conclusions are to be drawn about the relation between learning and insight. Insensitivity of Explicit Tests There are several other problems with the explicit tests commonly used to measure task knowledge and self-insight. First, these measures tend to be retrospective, asked after participants have completed numerous trials, and this can distort the validity of their assessments. The reliance on memory, possibly averaged across many trials, can make it difficult to recall a unique judgment strategy. This is especially problematic if people’s strategies have varied during the course of the task, making it hard if not impossible to summarize in one global response. In general it is better to get multiple subjective assessments as close as possible to the actual moments of judgment (cf. Ericsson & Simon, 1980; Harries & Harvey, 2000; Lovibond & Shanks, 2002). Second, explicit tests often require verbalization, but this can also underestimate task knowledge because participants may know what they are doing but be unable to put this into words. This is particularly likely in probabilistic tasks, where natural language may not be well adapted to the nuances of probabilistic inference. A third problem, prevalent in the neuropsychological literature, is that explicit tests are often too vague (Lovibond & Shanks, 2002). Rather than focus on specific task-relevant features they include general questions that are tangential to solving the task (e.g., questions about location of cards on screen). Once again this reduces the sensitivity of the test to measure people’s relevant knowledge or insight. This can lead to an over-estimation of Insight and strategy 7 insight (e.g., someone may be able recall features of the task irrelevant to good task performance) or to under-estimation (e.g., the questions fail to ask about critical information). The studies in this paper seek to improve the sensitivity of explicit tests on all these counts. Separate explicit tests will be used for task knowledge and self-insight, and both will involve specific questions of direct relevance to the task. In the case of task knowledge these will concern probability ratings about cue-outcome relations, in the case of self-insight subjective ratings about cue usage. Such ratings-based tests will avoid any problem of verbalization. To tackle the problem of retrospective assessments we will take multiple judgments during the course of the task (either blocked or trial-by-trial). Analyses of Judgment and Learning Claims about the dissociation between insight and learning also depend on an appropriate analysis of learning performance. In tasks with a dichotomous outcome, such as the WP task, learning is measured in terms of the proportion of correct predictions over a sequence of trials. Standard analyses then average across both individuals and trials to produce a mean percentage correct for the whole task. While this approach is useful for broad comparisons across different tasks, it provides no information about the dynamics of individual judgment and learning. The Lens Model A richer approach is provided by the lens model framework (for overviews see Cooksey, 1996; Goldstein, 2004), which assumes that people construct internal models to reflect the probabilistic structure of their environment. A central tenet is that judgmental processes should be modeled at the individual level before any conclusions can be drawn by averaging across individuals. An individual’s judgment policy is captured by computing a multiple linear regression of their judgments onto the cue values (across all task trials). The resultant coefficients for each cue are then interpreted as their subjective weights (or cue Insight and strategy 8 utilization weights). In a parallel fashion the objective weights for that individual are computed by a multiple regression from the outcomes experienced onto the cue values. This technique allows for the possibility that different individuals experience different objective weights. Individual learning is then assessed by comparing subjective and objective weights. Moreover, task knowledge can be assessed by comparing an individual’s objective weights with their explicit ratings of the cue-outcome associations, and self-insight assessed by comparing their subjective weights with their explicit ratings of their own cue usage. The lens model framework thus provides an analysis of individual judgment. However, although it avoids the loss of information incurred by averaging over participants, it still loses information by averaging over trials. It fails to capture the dynamics of a learning task – both in terms of potential changes in the environment, and potential changes in a judge’s policy. In particular, the reliance on global weights ignores the fact that both subjective and objective weights can vary across the course of a task. This problem arises even with stationary environments (as in the WP task), because the cue-outcome patterns experienced early on in the task might not be representative of the global structure. Analyzing individuals just in terms of their averaged performance across all the trials ignores this possibility, and can under-estimate performance. Moreover, it overlooks the possibility that individuals might change their judgment policies over time, and that such changes may track variations in the actual environment. A related shortcoming is that these global analyses assume that the judge has a perfect memory for all trials, and treat earlier trials in exactly the same way as later ones. But both of these assumptions are questionable – people may base their judgments on a limited window of trials, and may place more emphasis on recent trials (Slovic & Lichtenstein, 1971). Dynamic Models of Learning Insight and strategy 9 The need for dynamic models of learning is now widely recognized (Dayan, Kakade & Montague, 2000; Friedman, Massaro, Kitzis & Cohen, 1995; Kitzis et al., 1998; Smith et al., 2004). A natural extension to the lens model is the ‘rolling regression’ technique introduced by Kelley and Friedman (2002) to model individual learning in economic forecasting. In their study participants learned to forecast the value of a continuous criterion (the price of Orange juice futures) on the basis of two continuous-valued cues (local weather hazard and foreign supply). Individual learning curves were constructed by computing a series of regressions (from forecasts to cues) across a moving window of consecutive trials. For example, for a window size of 160 trials, the first regression is computed for trials 1 to 160, the next for trials 2 to 161, and so on. This generates trial-by-trial estimates (from trial 160 onwards) for an individual’s cue utilization weights, and thus provides a dynamic profile of the individual’s learning (after trial 160). Each learning profile was then compared with the profile of an ‘ideal’ learner exposed to the same moving window. Regressions for each ideal learner are also computed repeatedly for a moving window of trials, but in this case the actual criterion values (prices) are regressed onto the cues. The estimates of the ideal learner thus correspond to the best possible estimates of the objective cue weights for each window of trials. This technique provides dynamic models of both actual and ideal learners, and permits comparisons between them over time. For example, Kelley and Friedman showed that whereas ideal learners converged quickly on the actual weights, the participants learned more slowly, and over-estimated these weights towards the end of the task. In this paper we will use similar techniques to analyze individual learning, but with a smaller window size to simulate a more realistic memory constraint. This should provide a finer-grained analysis of the dynamics of individual learning. Insight and strategy 10 It is important to note that even an ideal learner is not guaranteed perfect performance on a probabilistic task. First, the noisy nature of the environment means that optimal performance is bounded. In the WP task 83% correct predictions is the best that can be expected even if the learner knows the objectives probabilities for each cue pattern. Second, predictive success also depends on the learner’s choice rule (Friedman & Massaro, 1998; Nosofsky & Zaki, 1998). Given their estimates for the outcome probabilities (conditional on each cue pattern) they might always select the most probable outcome (maximize) or distribute their choices to match the outcome probabilities (matching). Only the former leads to optimal performance. This issue will be revisited in the General Discussion. Strategy Analyses An alternative approach is developed by Gluck et al. (2002). They identified three main strategies employed in the WP task: (1) a multi-cue strategy that used all cards to predict the outcome; (2) a singleton strategy that only used cue patterns with a single card (and guessed on all other patterns); and (3) a one-cue strategy that used just one card (and ignored any other cards present). Ideal profiles were constructed for each of these strategies, and these were fit to the participant data. Overall the best fitting model was the singleton strategy, with a shift from singleton to multicue strategies as the task progressed. This improves over previous analyses of the WP task, and resonates with recent work on the use of simple heuristics (Gigerenzer et al., 1999). However, it suffers from the use of global cue-outcome associations to construct the ideal profiles for each strategy. This ignores the possibility that participants encounter a non-representative sampling of the environment early on in the task, and therefore may under-estimate the number using a multi-cue strategy. For example, in fitting the multi-cue strategy it is assumed that participants know the correct cue-outcome associations from the outset. This is an unrealistic assumption. Insight and strategy 11 The rolling regression method overcomes this problem by comparing individual learners to an ideal learner exposed to the same information. It also offers a more parsimonious explanation of the apparent strategy shifts in Gluck et al. (2002). People appear to use sub-optimal strategies early on in the task because they are in the process of learning each of the cue weights. The apparent shift to multi-cue strategies emerges as they learn all of these cue weights. The existence of both strong and weak cues will promote this pattern: stronger cues will tend to be learned before weaker ones. A single learning mechanism thus suffices to explain the experimental data. Strategy and Insight What is the relation between strategy and insight? Gluck et al. (2002) assume that strategies are implicit, so there need be no connection between the two. In contrast, opponents of the implicit/explicit distinction would maintain that the two are closely related. For example, multi-cue strategies require richer explicit knowledge than single-cue strategies. By using better and more frequent measures of explicit knowledge, and correlating these with implicit measures of learning, this question can be addressed. In sum, this paper aims to extend our understanding of how people learn in multiple cue tasks, and re-examine the prevalent claim that good performance can be achieved in the absence of insight. In doing so, it will introduce more appropriate measures of both task knowledge and self-insight, and more fine-grained methods for analyzing the dynamics of individual learning.