Content Generation for Strategy Games via LSTM

ALEKSANDR W. FRITZ

Procedural content generation (PCG) has been around since the early days of video games. One of the earliest instances of PCG was the game Rogue, a game where levels, enemies, and rewards were procedurally generated each playthrough so as to keep the content fresh and evergreen. This spawned the Rogue-like genre, which, among other things, has PCG as one of its defining features. This genre inspired many games with complex algorithms seen today such as Dwarf Fortress, and Minecraft. These game mainly use hand crafted functions, though, and as such, they can take long periods of time to create and design on their own. They are also very controlled, which can eventually lead to stagnant and too similar content over long periods of time when the goal is to create evergreen content. To discover new ways to generate content, many researchers have been looking into procedural content generation
via machine learning techniques. Through this approach, level designers can directly design levels as they would like them to be. These levels are then fed into a machine learning algorithm to train the model to create new levels that are similar to what the designers intended. The levels need not be too similar, though, and research has been proposed to attempt to take designers inputs and output novel and previously unthought of level designs and approaches. In this project, we apply long short-term memory (LSTM) network to the strategy game Wargroove for map generation. Wargroove is a turn based strategy game where the objective is to defeat the opponent’s commander or stronghold. It is played on a discrete two dimensional orthogonal grid. All distances are calculated in number of tiles on the grid. Grid tiles can only contain one unit and one map feature with some minor exceptions. This discrete approach with a strictly defined ruleset means that transferring maps into simple data structures is easy and lossless.Moreover, this project extends previous work on map generation to both map generation and player armies generation together and, not separately.