Skeleton-based multi-agent opponent search
Wael Al Enezi, Clark Verbrugge · 2021 IEEE Conference on Games (CoG) · 2021
In many games, players may run away and hide from NPC enemies that have previously observed them, either to avoid combat or as part of pursuing a stealth-based solution. Rational NPC response then requires searching for the hidden player, which for maximal realism should build on the last known location, and consider the relative likelihood of a player hiding or reaching each searched location. Unfortunately, search behavior is not usually systematic, and in practice is either limited to randomized goals within a small region, or exploits global information on the player position that should be unknown. In this work, we introduce a real-time method for directing a multi-agent search utilizing the environment's topology. This approach allows for more natural and wider-scoped search behavior. Experimental results show that this method scales to relatively large game maps, and performs better than or close to a naïve team of agents fully aware of the player's position.