Planning for millions of NPCs in Real-Time

We address the problem of scaling the generation of plans in real-time to control the behaviors of several millions of Non-Player Characters (NPCs) in video-games and virtual worlds. Search-based action planning, introduced in the game F.E.A.R. in 2005, has an exponential time complexity managing at most several tens of NPCs per frame. A close study of the plans generated in first-person shooters shows that: (1) states are vectors of enumerated values, (2) both initial and final states can be totally defined, (3) actions are both post-unique and unary, (4) plans are totally ordered, and (5) actions occur only once in plans. (1) to (5) satisfy the Simplified Action Structure (SAS) linear time planning framework SAS-PUT1. We strengthen previous claims on this framework saying that the associated linear time algorithm P is capable of managing several millions of NPCs per frame by testing it on three new realistic benchmarks that are based on commercial video games.