Information foraging with an oracle

Abstract

During ecological decisions, such as when foraging for food or selecting a weekend activity, we often have to balance the costs and benefits of exploiting known options versus exploring novel ones. Here, we ask how individuals address such cost-benefit tradeoffs during tasks in which we can either explore by ourselves or seek external advice from an oracle (e.g., a domain expert or recommendation system). To answer this question, we designed two studies in which participants chose between inquiring (at a cost) for expert advice from an oracle, or to search for options without guidance, under manipulations affecting the optimal choice. We found that participants showed a greater propensity to seek expert advice when it was instrumental to increase payoff (study A), and when it reduced choice uncertainty, above and beyond payoff maximization (study B). This latter result was especially apparent in participants with greater trait-level intolerance of uncertainty. Taken together, these results suggest that we seek expert advice for both economic goals (i.e., payoff maximization) and epistemic goals (i.e., uncertainty minimization) and that our decisions to ask or not ask for advice are sensitive to cost-benefit tradeoffs.

Fig 1. Experiment design.

We performed two studies (Study A and B), both using a within subjects, 2x2x2x2 design. We manipulated strategy choice, map size, strategy gap size and oracle accuracy. The gem layout (randomized for Study A, and spatial for study B) affected only the strategy choice factor—that is, the effect of the two primary strategies: call oracle first, or never call—shown in red and blue text above. See the main text for details.

Fig 2. Task screen layout.

See the main text for explanation.

Fig 3. Oracle (under)-use against optimal strategy.

Use was very low on NC trials, but only 55.8% of participants correctly used the oracle on OF trials. An interaction effect is seen showing that larger gaps were associated with increased use of the oracle on OF trials ($F(1,27)=35.9,p<.001,{\eta }_p^2=.066$).

Fig 4. Scatter plots of oracle requests vs mean score across participants.

Mean score is defined as the mean of total points collected by a participant across all non-practice trials. Total Oracle Requests is the count of trials where the participant called the Oracle, across all non-practice trials. A significant positive correlation is seen indicating the relationship between correctly identifying problems where the oracle is useful, and overall performance. Correlation is significant with Pearson-r = 0.78, p < 0.0001.

Fig 5. Strategy satisfaction, as measured by percent of participants responding yes to question: “If you could retry the same problem again, would you use the same strategy to call the oracle [not call the oracle?]”, post-trial.

Here we analyze trials where the oracle was accurate, to simplify analysis of optimal behavior. We find a main effect showing higher satisfaction when the oracle was used, whether optimally or non-optimally ($F(1,2)=21.7,p=.043,{\eta }_g^2=.684$). Interestingly, three quarters (74.5%) of participants said they’d keep the same strategy even when they used a non-optimal strategy.

Fig 6. Comparison of trial score across participants who did not call the oracle (transparent blue points, with mean shown by opaque blue dot) versus trial’s expected score under a random policy (black dot).

Mean performance was above expectation for most trials indicating that spatial information was exploited when participants chose not to call the oracle. Maps are ranked by expected score (“Gem EV Rank”).

Fig 7. A positive correlation (r = 0.35, p = 0.06) is observed between participants’ IUS score and use of the oracle on VT trials.

Fig 8. Use of the oracle vs IUS types (under or over median IUS score) in the spatial group (Study B).

Left: among variance trade-off trials, right: among never call optimal trials. More requests were seen among high-IUS participants ($F(1,27)=6.1,p=.019,{\eta }_p^2=.115$), and in VT trials ($F(1,27)=158,p<.001,{\eta }_p^2=.714$). We additionally find an interaction between IUS-type and trial optimal strategy ($F(1,27)=9.2,p=.005,{\eta }_p^2=.127$), showing that the most oracle requests were seen among high-IUS participants in VT trials (T = 3.05, p = .005, BF = 8.6).

Fig 9. Oracle use trends through experiment (VT trials only, spatial group).

Lines show a rolling average of oracle use across a 5-trial window, for high-IUS participants (blue line), low-IUS participants (black line), and across all participants (gray line).