Neural encoding of perceived patch value during competitive and hazardous virtual foraging

The study investigates how humans make foraging-like decisions when facing varying levels of competition and predation threat. Drawing on optimal foraging theory and the selfish herd/risk-dilution framework, the authors posit that perceived patch value (PPV)—a function of reward availability, competition density, and predation threat—governs patch choice. In safe contexts, avoiding competition should maximize rewards; under threat, congregating with others should dilute predation risk. The research questions are whether humans adapt their decisions according to PPV across safe versus threat contexts and whether PPV, rather than observable social density alone, is encoded neurally. The authors hypothesize context-dependent shifts in strategy (competition avoidance in safety, risk dilution under threat) and neural encoding of PPV independent of social density, particularly in MCC and vmPFC.

Prior foraging research highlights the influence of competitor density, food quantity, and energy costs on decisions. In the absence of predation, avoiding competitor-dense patches maximizes gains; under predation risk, larger groups reduce individual capture probability (risk dilution), albeit with efficiency costs. Human and non-human primate studies using virtual foraging tasks without threat implicate MCC, vmPFC, and dorsal anterior cingulate (dACC) in action selection, value encoding, and decision difficulty. vmPFC is linked to choice value and goal-directed valuation; dACC findings relate to foraging decision difficulty, though debated. The authors situate their work by introducing predation threat into a two-patch paradigm to test whether PPV, not social density per se, is the key neural decision variable.

Participants: 22 healthy adults (6 female; mean age 31, range 18–49) completed a virtual two-patch foraging task across two days (about 4 hours scanning total). One session’s data were lost for one participant and neural data for another, yielding n=20 for behavioral and neuroimaging analyses (n=19 for RSA after exclusion based on condition exposure). Threat trials included the possibility of electric shock to the left wrist, calibrated individually to an aversive yet tolerable level. IRB-approved; informed consent obtained.

Task design: Two side-by-side patches contained 1–6 AI competitors. Participants observed the competitor configuration (3 s), then had 3 s to choose a patch. In safe trials, they foraged for tokens; in threat trials, a predator could appear randomly and capture one agent at random, ending the trial and delivering a shock if the participant was captured. Reward tokens appeared at random times/locations; AI competitors moved at the same speed and obstructed movement. Patch switches had no cost; safe and threat patches were matched in effort, energy costs, competition, and reward density.

Cycles: Each block used fixed, repeating cycles of competitor states per patch (three-state cycles), enabling prediction of upcoming configurations. Four blocks across two days employed different repeating sequences.

Trial types: Short immediate (SI), short later (SL), and long immediate (LI). SL allowed choosing the next cycle state; LI lasted 14 s with a state change at 7 s but remained within the same context (safe or threat). Later decisions were rare (7.6%); analyses focused on immediate decisions (SI, LI). Each block contained eight sub-blocks of 18 trials (balanced safe/threat), totaling 576 trials.

Behavioral data: x–y coordinates (30 Hz), reward events, patch choices, competition levels, captures/shocks were recorded. Analyses examined choice tendencies by condition, trial length, block, and competitor numbers.

Behavioral modeling (PPV): For each participant, PPV for each condition (combination of competitor number and threat level) was computed as the average accumulated points minus expected shock costs: PPV = points_X + P(shock)X × shock_cost, with shock_cost a free parameter (negative for all but two participants). Choice probabilities were generated via softmax with temperature τ using only displayed patch options. Model comparison included: (1) a Bayesian mean tracker learning model updating expected value per condition with learning rate G_t = v{t-1}/(v_{t-1}+θ), θ=1; (2) a stickiness variant biasing repetition of the previously chosen patch. Models were fit in PyMC3 using variational inference; WAIC was used for comparison. The initial PPV model fit best (WAIC 6181.56) versus learning (WAIC 6219.19) and stickiness (WAIC 9829.30).

fMRI acquisition: 3T Siemens Prisma, 32-channel head coil. T2* EPI: TR/TE=1000/30 ms, flip angle 60°, 72 slices, multiband 6, 2.0 mm isotropic, no in-plane acceleration, 3/4 partial Fourier, FOV 192×192 mm, matrix 96×96, slice angulation 20°. Anatomicals: 0.9 mm T1w MEMP-RAGE and 3D T2w SPACE.

Preprocessing: fMRIPrep pipeline: N4Bias correction; ANTs-based skull strip and normalization to ICBM 152 (2009c); FAST tissue segmentation; motion correction (MCFLIRT); BBR co-registration; combined transforms via ANTs; CompCor (6 aCompCor, 6 tCompCor); framewise displacement computed. Visualization/processing via Nilearn.

Representational Similarity Analysis (RSA): First-level GLMs (FSL 6.0) modeled each decision period as a separate regressor; other periods and shock modeled as regressors of no interest; nuisance regressors included motion, FD, WM/CSF. Beta maps per trial were used for whole-brain searchlight RSA (6 mm radius; 3 mm FWHM smoothing). Neural RDMs (Spearman correlation distance across trials) were regressed onto task RDMs representing: PPV of current patch, PPV of alternative patch, PPV difference, number of competitors (current, alternative, difference), threat level, and run/session similarity. Linear regression within each searchlight yielded β maps per RDM. Group statistics used FSL randomize with TFCE (5000 permutations, 10 mm variance smoothing), thresholded at p<0.05 TFCE-corrected.

Univariate analyses: Data smoothed 8 mm FWHM; decision period modeled with parametric modulators: PPV and competitor number (current/alternative), their differences, and threat level. Second-level fixed effects within participant; group-level randomize with 5000 permutations and 10 mm variance smoothing. Additional GLMs examined chosen vs. unchosen value components.

Behavior: Participants adopted context-appropriate strategies. They chose the less-populated patch in 89% of safe decisions and in 32% of threat decisions (χ²=4046, p<0.0005, 95% CI [0.55, 0.58]; paired t(20)=9.50, p<0.0005, 95% CI [0.44, 0.69], mean difference=0.57). Difference scores reflecting competitor imbalances differed by context (paired t(20)=9.66, p<0.0005, 95% CI [2.53, 3.93]). In threat, selecting patches with more competitors reduced capture probability via risk dilution across difference-score bins (all non-paired t-tests between positive and negative values p<0.005). Participants collected fewer tokens in threat than safe (paired t(20)=13.11, p<0.0005, 95% CI [34.64, 47.75]).

Behavioral modeling of PPV: The PPV model captured the context-dependent effects of competition (PPV decreased with more competitors in safety, increased with more competitors under threat). The PPV difference between patches predicted choices, correctly classifying 69.89% of choices, with well-calibrated predicted probabilities. Model comparison favored the initial PPV model (WAIC 6181.56) over learning (WAIC 6219.19) and stickiness (WAIC 9829.30).

Neural RSA: PPV of the alternative patch was encoded in MCC, PCC, medial prefrontal cortex (mPFC), and orbitofrontal cortex (OFC) (TFCE p<0.05). PPV of the current patch was encoded in overlapping regions (MCC, PCC, mPFC), and additionally in premotor cortex, hippocampus, and anterior insula. These regions encoded PPV but not social density per se; no areas represented the number of competitors in current or alternative patches, nor the PPV or social-density difference between patches. Amygdala did not significantly encode PPV (trend for current patch only). Threat level (safe vs. threat) was represented broadly, including MCC, vmPFC, hippocampus, and amygdala.

Univariate: More competitors in the current patch increased visual cortex activity; more competitors in the alternative patch reduced visual cortex activity. PPV effects: higher current-patch PPV related to greater thalamic activity; higher alternative-patch PPV associated with greater OFC/mPFC activity and reduced thalamus, insula, dorsal striatum activity. No significant univariate clusters for threat during the decision phase. Chosen–unchosen value difference showed widespread negative effects, including dACC and dorsolateral PFC; absolute value difference associated with a right DLPFC cluster; chosen/unchosen values mainly in occipital areas.

Findings show humans flexibly adjust foraging decisions according to perceived patch value: they avoid competition to maximize rewards when safe and seek larger groups to dilute predation risk under threat. PPV, integrating reward, competition, and threat, better explains behavior than social density alone. Neurally, PPV is represented in a distributed network centered on vmPFC, OFC, MCC, and PCC during decision evaluation. The MCC appears to represent both current and alternative option values and may coordinate affective/motor responses based on learned values, particularly under threat, while vmPFC represents option values to support decision making. Unlike some prior work emphasizing difference signals or dACC involvement in difficulty, here PPV representations were independent for each option and dACC effects were not prominent, likely due to matched difficulty across conditions. Multivariate RSA revealed representational structure consistent with PPV rather than simple activity-level changes, distinguishing this work from prior univariate-only approaches.

The study introduces and formalizes perceived patch value (PPV) as a unifying decision variable that captures competition avoidance in safety and risk-dilution under threat. Behaviorally, PPV predicts patch choices across contexts; neurally, PPV is encoded in vmPFC and MCC among a broader network, independent of social density or simple value differences. These findings suggest domain-general valuation and action-integration mechanisms support adaptive foraging decisions in social environments with predatory risk. Future work should dissect how PPV is computed (model-free vs. integrative online computation), incorporate measures such as decision confidence and reaction time, and explore learning and uncertainty in dynamic environments.

The study cannot definitively determine how participants compute PPV; values could be learned model-free rather than computed online during decision making. The task design prioritized separating decision evaluation from action, limiting reaction time measures and assessment of decision confidence. The paradigm was not optimized to study learning or uncertainty; a learning model did not improve fit given the stable environment and extensive practice/lengthy task. Multivariate RSA identifies where variables are represented but not the coding scheme; potential confounds like subjective value and confidence may overlap with PPV. dACC involvement may be minimized due to matched decision difficulty across conditions.