Joint commitment in human cooperative hunting through an “Imagined We”

Cooperation involves the challenge of jointly selecting one from multiple goals while maintaining the team’s joint commitment to it. We test joint commitment in a multi-player hunting game, combining psychophysics and computational modeling. Joint commitment is modeled through an "Imagined We" (IW) approach, where each agent uses Bayesian inference to infer the intention of “We”, an imagined supraindividual agent controlling all agents as its body parts. This is compared against a Reward Sharing (RS) model, which frames cooperation through reward sharing via multi-agent reinforcement learning (MARL). Both humans and IW, but not RS, maintained high performance by jointly committing to a single prey, regardless of prey quantity or speed. Human observers rated all hunters in both human and IW teams as making high contributions to the catch, regardless of their proximity to the prey, suggesting that high-quality hunting stemmed from sophisticated cooperation rather than individual strategies. Unlike RS hunters, IW hunters are capable of cooperating not only with one another, but also with human participants actively engaged in the same hunting game. In conclusion, this study demonstrates that humans achieve cooperation through joint commitment that enforces a single goal, rather than simply motivating members through reward sharing.

Methods
Data Collection:
The dataset was collected through offline laboratory experiments involving human subjects, machine simulations, and human-machine collaboration. The machine simulations and collaborative tasks were based on models implemented in Python, including reinforcement learning, neural networks, and Bayesian inference. Participants engaged in tasks designed to assess cognitive processes, with data recorded in controlled laboratory conditions.

Data Processing:
The collected data was processed using Python’s pandas library. This involved data cleaning, transformation, and preparation for further analysis. For statistical analysis, we used the Jeffreys’s Amazing Statistics Program (JASP) software to perform various statistical tests.