A rtificial intelligence (AI) is one of the most disruptive technologies of our generation. Media coverage of this disruption tends to focus on the more contentious ways in which organizations could use AI to get ahead. International concern has strengthened this focus: in July 2018, more than 2,000 researchers from 36 countries signed a pledge not to develop lethal autonomous weapons. In the same year, Google instigated a ban on developing AI software that could be used to make weapons. The move came after several employees resigned over the company's contract with the US Department of Defense to analyze remotely piloted aircraft videos. But this attention on autonomous weapons detracts from the main benefit AI brings to defense organizations: the ability to analyze data too vast for humans.

Artificial intelligence (AI) is pervading nearly in every sphere of our lives. Banks, hospitals, hotels, manufacturing, and other industries are actively deploying AI-based solutions for efficiency, cost savings, and higher profit. The defense industry is the latest sector to utilize AI. Recently, a group of 116 leading AI experts wrote an open letter to ban AI use for developing autonomous weapons. The experts are calling use of AI as the third revolution in warfare after gunpowder and nuclear weapons.

Game tree search algorithms, such as Monte Carlo Tree Search (MCTS), require access to a forward model (or "simulator") of the game at hand. However, in some games such forward model is not readily available. In this paper we address the problem of automatically learning forward models (more specifically, combats models) for two-player attrition games. We report experiments comparing several approaches to learn such combat model from replay data to models generated by hand. We use StarCraft, a Real-Time Strategy (RTS) game, as our application domain. Specifically, we use a large collection of already collected replays, and focus on learning a combat model for tactical combats.