Agents Based on Program Trees and Genetic Algorithms

Agents solve unknown logistical problems with different strategies. In the AbstractSwarm Multi-Agent Logistics Competition (2024), different strategies are compared by performance. Agents chose their next action based on a value for it. The program trees presented here represent an arithmetic expression that combines different base factors to define the next action. Combinations are based on probability and modified during a simulation. The tree structure supports the creation of new arithmetic expressions and is simple to extend or modify. The presented genetic algorithm is based on these trees and extends the limited number of available expressions. An available tree is manipulated with imitations of mutation and recombination to create any combination of base factors. More possible combinations cover a larger search space but require more engineering work. Both of the presented strategies were implemented for the AbstractSwarm Multi-Agent Logistic Competition (2024). An earlier version of the tree variant (DecisionTreeOptimizingAgent) reached second place with a score \({\approx }36.81\) (first: \({\approx }35.98\) ). The further developed genetic algorithm which is based on the program tree algorithm might achieve even better results but has to be further tested.