A Cognitive Neuroscience Study in Multi-Agent Box-Pushing Adversarial Games

A co-evolutionary spiking neural network was tested in a 20×10 grid box-pushing adversarial game.

After evolutionary training, 1000-step behavioral data were recorded under fixed weights and R-STDP.

With fixed weights, the system logged 42 pushes, 10 attacks, 10 rescues, and 6.10% exploration; the 10 rescues included 9 counter-kills and 1 teammate counter-attack.

Fixed weights also kept communication constant and non-functional, with bimodally polarized brain network weights and team scores of about 500 vs 0.

Under R-STDP, the system logged 14 pushes, 14 attacks, 11 rescues, and 24.55% exploration; pulses tracked energy (r=0.56), weights stayed bimodally polarized, and the left team scored about 80 vs 0.