Human-Inspired GNP lets AI agents evolve like children, boosting performance

A team led by Ali Kohan from multiple institutions introduced Human-Inspired Genetic Network Programming (HGNP), a framework that dynamically regulates the exploration-exploitation trade-off in self-evolving agents. The approach draws from human developmental psychology: children prioritize broad experimentation and action over deliberation, a tendency that reverses with age. By mapping this to GNP's judgment nodes (deliberation) and processing nodes (action), HGNP applies novel adaptive crossover and mutation operators, plus a cycle elimination mechanism. This allows the agent to explore more early in evolution and exploit later, without manual tuning of probabilities.

The researchers integrated HGNP with standard GNP and two recent variants (SBGNP and another) and tested them on the Tileworld benchmark—a classic multi-agent environment requiring planning under uncertainty. The HGNP-SBGNP combination delivered the best strategy performance, significantly outperforming baseline methods. Because the modifications are general, HGNP can be applied to almost all GNP variants, offering a plug-and-play upgrade for autonomous agent design. This work advances explainable AI by making evolutionary graph-based methods more adaptive and human-like.