Modeling Human Behavior in a Strategic Network Game with Complex Group Dynamics

Human networks greatly impact important societal outcomes, including wealth and health inequality, poverty, and bullying. As such, understanding human networks is critical to learning how to promote favorable societal outcomes. As a step toward better understanding human networks, we compare and contrast several methods for learning models of human behavior in a strategic network game called the Junior High Game (JHG) [39]. These modeling methods differ with respect to the assumptions they use to parameterize human behavior (behavior matching vs. community-aware behavior) and the moments they model (mean vs. distribution). Results show that the highest-performing method, called hCAB, models the distribution of human behavior rather than the mean and assumes humans use community-aware behavior rather than behavior matching. When applied to small societies, the hCAB model closely mirrors the population dynamics of human groups (with notable differences). Additionally, in a user study, human participants had difficulty distinguishing hCAB agents from other humans, thus illustrating that the hCAB model also produces plausible (individual) behavior in this strategic network game.

💡 Research Summary

The paper tackles the problem of modeling human decision‑making in a complex strategic network environment by using the Junior High Game (JHG) as a testbed. In JHG each player receives a fixed number of tokens each round and must decide how many to keep, how many to give to other players, and how many to use for attacks. The effect of each token on a player’s “popularity” is weighted by the giver’s current popularity, creating a dynamic, directed, weighted, signed network that evolves over many rounds. This setting captures many real‑world social processes—resource exchange, alliance formation, power asymmetry, and collective retaliation—far beyond the binary cooperate/defect dilemmas typical of classic social‑dilemma games.

The authors propose four modeling configurations that differ along two orthogonal dimensions. The first dimension concerns the behavioral assumption used to parameterise agents: (1) Behavior Matching (TFT), which assumes agents simply reciprocate the actions they received in the previous round, and (2) Community‑Aware Behavior (CAB), which assumes agents form, join, and defend groups, allocating tokens to support in‑group members and to attack out‑group members. The second dimension concerns the learning objective: (a) Particle Swarm Optimization (PSO), which searches for a single parameter set that minimises the average error between a model’s token allocations and the observed human allocations (i.e., modelling the mean of the human population), and (b) Evolutionary Population Distribution Modeling (EPDM), a novel evolutionary algorithm that simultaneously learns a distribution of 100 parameter sets, each representing a plausible individual strategy within the human population. EPDM uses a custom similarity score that evaluates positive, negative, and keep‑token components as well as a penalty for contradictory actions, thereby capturing richer behavioural nuances than simple mean‑squared error.

The experimental pipeline consists of (i) collecting a few hundred human games of JHG, (ii) fitting each of the four configurations (TFT‑PSO, TFT‑EPDM, CAB‑PSO, CAB‑EPDM) to this data, (iii) deploying the resulting agents in small simulated societies and comparing macro‑level dynamics (popularity trajectories, token flow, group formation/dissolution) to those observed in the human data, and (iv) conducting a user study in which participants interact with agents and are asked to identify which players are human.

Results show that the CAB‑EPDM configuration—referred to as hCAB—outperforms all others on both macro and micro metrics. hCAB agents reproduce the population‑level popularity curves and the emergence of alliances and rivalries seen in real players, while also generating individual token‑allocation patterns that human observers could not reliably distinguish from genuine human behaviour. In contrast, TFT‑based agents tend to over‑react with retaliation, leading to unstable popularity dynamics and easy detection by participants. The PSO‑derived mean models, even when using the CAB behavioural assumption, capture only the average trend and miss the variability necessary for realistic group‑level outcomes.

The paper’s contributions are threefold. First, it clearly delineates two competing cognitive hypotheses for human strategic interaction (reciprocity vs. group‑based reasoning) and demonstrates how these hypotheses manifest in a richly structured game. Second, it introduces EPDM, an evolutionary approach that learns a distribution of strategies rather than a single average, addressing the common limitation of data‑scarce domains where collecting massive behavioural datasets is infeasible. Third, it provides empirical evidence that a distribution‑aware, community‑aware model (hCAB) can both match observed population dynamics and generate plausible individual behaviour, suggesting that future simulations of social systems—whether for policy testing, educational tools, or AI‑human interaction research—should prioritize modelling behavioural heterogeneity and group‑level cognition.

Overall, the study advances our understanding of how to construct interpretable, data‑efficient models of human behaviour in networked strategic settings, and it offers a concrete, validated framework (hCAB) that can be adapted to other games or real‑world social platforms where group dynamics and resource competition play a central role.