Abstract
There is an ongoing discussion in the game AI community, whether to use AI planning for controlling non-player characters (NPCs) in computer games. Recent years have seen implementations of both state space and hierarchical task network (HTN) planning in AAA game titles, each having specific advantages and disadvantages.
This paper is concerned with the performance aspect of both technologies and proposes a general methodology for comparing performance of action selection mechanisms. Two case studies comparing NPCs controlled by the JSHOP2 HTN planner and SGPlan 6 and Metric-FF state space planners in a game-like competitive environment are presented and their results are statistically analyzed.
First environment is a puzzle-like scenario with minor combat element and strong competition for limited resources. Second environment features two pairs of cooperating agents in a combat scenario inspired by the Gears of War series.
It is concluded that in puzzle-like domains HTN planning is inferior to state space planning, but for the second domain, the results are not conclusive.