Heuristics for sleep and heal in combat

Abstract

Basic attack and defense actions in games are often extended by more powerful actions, including the ability to temporarily incapacitate an enemy through sleep or stun, the ability to restore health through healing, and others. Use of these abilities can have a dramatic impact on combat outcome, and so is typically strongly limited. This implies a non-trivial decision process, and for an AI to effectively use these actions it must consider the potential benefit, opportunity cost, and the complexity of choosing an appropriate target. In this work we develop a formal model to explore optimized use of sleep and heal in small-scale combat scenarios. We consider different heuristics that can guide the use of such actions; experimental work based on Pokémon combats shows that significant improvements are possible over the basic, greedy strategies commonly employed by AI agents. Our work allows for better performance by companion and enemy AIs, and also gives guidance to game designers looking to incorporate advanced combat actions without overly unbalancing combat.