Conservative forgetful scholars: How people learn causal structure through sequences of interventions.

Abstract

Interacting with a system is key to uncovering its causal structure. A computational framework for interventional causal learning has been developed over the last decade, but how real causal learners might achieve or approximate the computations entailed by this framework is still poorly understood. Here we describe an interactive computer task in which participants were incentivized to learn the structure of probabilistic causal systems through free selection of multiple interventions. We develop models of participants' intervention choices and online structure judgments, using expected utility gain, probability gain, and information gain and introducing plausible memory and processing constraints. We find that successful participants are best described by a model that acts to maximize information (rather than expected score or probability of being correct); that forgets much of the evidence received in earlier trials; but that mitigates this by being conservative, preferring structures consistent with earlier stated beliefs. We explore 2 heuristics that partly explain how participants might be approximating these models without explicitly representing or updating a hypothesis space.

Strategies to intervene on causal systems are adaptively selected

“…The information gain model (IG) in particular has been advanced as a normative strategy for a wide number of information search tasks (e.g., Austerweil & Griffiths, 2011;Gureckis & Markant, 2009;Klayman & Ha, 1987;Lindley, 1956;Markant & Gureckis, 2012a, 2012bNajemnik & Geisler, 2005;Nelson, Divjak, Gudmundsdottir, Martignon, & Meder, 2014;Oaksford & Chater, 1994, 1994, and was first applied to modeling causal interventions in the machine-learning literature (Murphy, 2001;Tong & Koller, 2001). It has also had an important influence on the psychology of learning from causal interventions (Bramley et al, 2014;Shafto et al, 2014;Steyvers et al, 2003). According to IG, interventions are made with the goal of decreasing the learner's uncertainty about a causal system, given a range of possible structures that explain its behavior.…”

“…For example, one proposal is that people search for information that can discriminate between possible hypotheses about causal structure, for instance by using an information gain (IG) strategy (Bramley, Lagnado, & Speekenbrink, 2014;Nelson, 2005;Shafto, Goodman, & Griffiths, 2014;Steyvers, Tenenbaum, Wagenmakers, & Blum, 2003). Alternatively, in the broader hypothesis testing literature many studies argue that people seek information that yields positive evidence to confirm a single hypothesis, disregarding alternatives (e.g., Klayman & Ha, 1987;Nickerson, 1998;Wason, 1960).…”

Testing one or multiple: How beliefs about sparsity affect causal experimentation

“…It also makes predictions for when a learner should stop making tests and guess the answer even if entropy is still nonzero. Evidence suggests that people do not typically plan multiple steps into the future when collecting information in causal systems (e.g., Bramley et al, 2015). We therefore chose to focus on the EIG analysis for the main body of this paper.…”

“…Compared to the deterministic case used in this paper, probabilistic causal links can make intervention-planning much more difficult (e.g. Bramley et al, 2015;Steyvers et al, 2003).…”

“…Furthermore, people are sometimes able to come up with highly efficient experiments that optimize information gained per intervention or minimize the total number of tests needed on average to discover the true causal structure (Bramley, Lagnado, & Speekenbrink, 2015;Bramley, Dayan, Griffiths, & Lagnado, 2017; Coenen, Rehder, & Gureckis, 2015;Steyvers, Tenenbaum, Wagenmakers, & Blum, 2003).…”