Genetic Algorithms for optimizing behavior trees in air combat

Abstract: Modelling and simulating entities in virtual environments are tools commonly used by companies to test, validate and verify their products in close to real scenarios; effectivelyreducing the cost, time and effort compared to real life testing. This is especially the case in the area of air combat where realistic behaviors are not only a necessity, but paramount to replace the costs of fuel and operation time. The behavior tree framework is a behavior model whichrepresents entity actions with regards to its perception of the world whilst being easy to manuallyvalidate through its intuitively structured nature. However, as different simulated scenarios require different behaviors, operators commonly has to manually craft new behavior trees at the cost of time and effort. In this thesis, the AI technique Genetic Algorithms (GA) is used to improve a previously crafted general behavior tree with regards to a given 4v4 beyond-visual-range air combat scenario. To this end, a select number of parameters within the behavior tree are optimized in two experiments where a) all parameters are optimized globally and b) the parameters are divided into blocks of sub-behaviors (Engage, Fire missile, etc.) which are then optimized individuallyand combined at a later stage. The agents in the GA are put against the base tree where the baseline is referred to as the base tree vs itself. As the problem proved too easy and resulted in an over-optimized behavior when a single scenario was used, the decision was made to increase the number of the scenarios to three; differing in positions and orientations. The former experimentresulted in a behavior capable of defeating all entities in the other team without any casualties in all three scenarios while the behavior in the latter experiment failed to find the cross-blockrelations, and thus, only achieved a slightly better result than that of the baseline. However, the parameters of highest importance are found to be highly correlated in both experiments and GA is concluded to be a satisfactory technique for the problem of generating improved behaviors with regards to given scenarios.