Portfolio greedy search and simulation for large-scale combat in starcraft

Abstract

Abstract-Real-time strategy video games have proven to be a very challenging area for applications of artificial intelligence research. With their vast state and action spaces and real-time constraints, existing AI solutions have been shown to be too slow, or only able to be applied to small problem sets, while human players still dominate RTS AI systems. This paper makes three contributions to advancing the state of AI for popular commercial RTS game combat, which can consist of battles of dozens of units. I. I BMulti-agent planning is an important sub-field of Artificial Intelligence research with many real-world applications that deal with agent cooperation, such as robotic search and rescue missions and unit coordination on a battlefield. Finding optimal actions for collections of agents if computationally intractable for all but the simplest planning tasks. For instance, finding shortest solutions in the generalized sliding tile puzzlewhich can be regarded as a restricted multi-agent path-planning problem -is NP-hard [1]. When constructing planning systems involving multiple agents acting in the world, we therefore have to resort to approximations which hopefully generate useful results, even under real-time constraints.Video games are fruitful application areas for multi-agent planning research. Often fast paced and featuring numerous game objects that can act independently, video games pose challenging research questions, such as: 1) How to quickly navigate groups of units on large maps, while staying in formation? 2) How to deal with imperfect information, such as unknown opponent locations? 3) How to cooperate well with team members? 4) How to detect and exploit opponent weaknesses? 5) How to coordinate attacks involving dozens of units?Most planning tasks in video games are time criticalany delay can be costly. Consider, for instance, a combat unit targeting system. Finding most effective target sequences is PSPACE-hard [2]. So, if we actually took the time to solve non-trivial instances of this problem optimally while playing, we could see ourselves outplayed by a much faster scripted solution that employs simple strategies such as "attack the weakest opponent unit in weapon's range". It is therefore important to balance plan quality and planning time, and to strive for systems that are able to interleave planning with plan execution.In recent years, developing AI systems for video games has gained attention in the AI research community due to the challenging problems these games pose, combined with the fact that human players still outperform AI systems in this application area, as demonstrated -for exampleat the AIIDE 2012 workshop on Artificial Intelligence in Adversarial Real-Time Games [3], where a strong, but not quite world-championship level S C player defeated the best S C AI systems without any problem. S C by Blizzard Entertainment is a popular real-time strategy game in which players try to defeat opponents by gathering resources, producing fighting units, and destroying their buildings o...

Script- and cluster-based UCT for StarCraft

“…The algorithm is used for finding optimal decisions by using random sampling. Churchill and Buro [3] have successfully applied a variation of the UCT algorithm called UCT Considering Durations (UCTCD) to combats in StarCraft. The branching factor of the UCTCD is however extremely large as each unit under control can choose from multiple possible actions.…”

“…The branching factor of the UCTCD is however extremely large as each unit under control can choose from multiple possible actions. The UCTCD was shown to be beaten by a greedy search algorithm, called Portfolio Greedy Search [3], which searches for sequences of scripts instead of individual unit actions. The final sequence of scripts is used to assign actions to units.…”

Modified Adversarial Hierarchical Task Network Planning in Real-Time Strategy Games

“…The evaluation function is a variant of LTD2 [27,28] that not only considers the hit-points of a unit, but also its costs.…”

Online Evolution for Multi-action Adversarial Games

“…Rapid Action Value Estimation (RAVE) is another popular enhancement that has been shown to improve MCTS in Go [9]. Script-based approaches such as Portfolio Greedy Search [5] and Script-based UCT [11] deals with the large branching factor of real-time strategy games by exploring a search space of scripted behaviors instead of actions.…”