The study investigates how macaques balance information-seeking with reward-seeking when foraging in an experimentally ambiguous terrain. Using an Exploration Room platform, four male rhesus monkeys forage for hidden food within a dense grid of woodchip piles, designed to obscure reward locations. The setup aims to reveal whether information about nearby locations shapes spatial search decisions in a novel environment, mimicking challenging natural foraging amid uncertainty.
Experimental design and methods
Environment: A floor grid with 81–108 piles; 21 piles conceal rewards per session. The central cluster of filled piles contains three rewards, with outer piles containing one, while the remaining piles are empty.
Protocol: Each session starts with a locally centered, randomly chosen filled-pile cluster to create a localized abundance map. Monkeys forage until they lose interest or depart.
Subjects and data: Four rhesus monkeys (Vin, Luk, Hum, Nat) were tested with 5–7 sessions per animal. Sessions were video-recorded, and every pile search was labeled as a discrete event, yielding a time-stamped two-dimensional coordinate sequence.
Measures: Foraging paths were analyzed by computing Euclidean step sizes between consecutive searches. The study quantified the fraction of filled piles found (42–100%), the earliest first find (one search) and the latest (22 searches). Step-size distributions were assessed for patch-wise bimodality, Brownian, and Lévy-like heavy tails using Hartigan’s dip test and double-log plots.
Key results
Movement patterns: Step-size distributions were heavy-tailed and unimodal, with no evidence of patch-wise bimodality. On double-log plots, distributions aligned with a Lévy-like pattern, with slopes in the range roughly between -1 and -3, suggesting occasional long-distance moves amid dense local sampling.
Local responses to rewards: After encountering a filled pile, monkeys tended to choose nearer piles than after encountering an empty pile (Vin p=0.02; Luk p=0.006; Hum p<<0.0001; Nat p=0.001). Despite proximity effects, paths did not collapse into dwelling near rewards; exploration continued beyond the immediate vicinity.
Path structure: A convolutedness index indicated modest roaming tendencies pre-reward, without systematic shifts to dwelling post-reward, implying sustained exploratory behavior rather than a strict exploitation mode.
Information-reward-proximity integration: A kernel-based generative model represented the search as a weighted combination of three maps: information (ambiguity), reward, and proximity. After each search, the ambiguity kernel was updated (or removed when known), reward kernels updated upon finds, and proximity reflected current location. Fitting to 100 random data subsets (hundreds of pile searches per animal) reproduced observed statistics: information-dominated settings yielded Brownian-like walks, balanced weights produced Lévy-like walks, and reward-dominant weights shortened post-reward steps.
Individual differences: Model fitting revealed distinct strategies (e.g., Vin emphasized information more than Luk or Hum; Luk favored rewards). Kernel sizes indicated overlap among unsearched piles, with ambiguity kernels typically the largest.
Map structure and generalization
Localized versus scattered abundance maps: Information weighting increased on scattered maps, indicating heightened information-seeking when rewards are dispersed. Vin and Nat exhibited stronger information weighting than Luk or Hum across map types, suggesting individual differences in balancing information, reward, and proximity.
Limitations
The study employs a small sample (n=4) and an artificial grid, which may limit extrapolation to wild macaque foraging. Nonetheless, the controlled paradigm provides ecologically informed insight into how information seeking complements energy preservation and reward pursuit in ambiguous terrains.