Investigating small scale combat situations in real time strategy computer games

This thesis presents and analyses the problem of small scale combat (SSC) in real time strategy (RTS) computer games. An RTS game is a war simulator where several opposing factions battle in a virtual world. The problem of SSC appears when soldiers, called units, of opposing factions meet in this virtual world. In commercial RTS games SSC situations are handled by applying simple rules to each unit involved such as “attack the nearest enemy unit”. The result is far from optimal as several examples will show. Therefore, we will investigate other methods for handling SSC situations. The commercial computer game industry’s name for modules handling computer controlled characters is Game AI. In this thesis we investigate which methods and algorithms this concept covers. This investigation includes an overview of selected computer game genres, an introduction to what quality in computer games is, and an overview of how Game AI is handled in different genres. We will define the concept of consistent characters. On this basis we will discuss how consistency influences the quality of a computer game. We will argue that if units in an RTS game are to be considered consistent they ought to behave optimally regarding the rules of the game world. In this thesis we present a method for solving SSC based on rule-induced timestamped game trees. Due to the amount of information in the nodes we investigate a machine learning approach to derive a node rating function. We achieve a reduction of the number of nodes in the tree by applying a sequence of rules to each node, thereby reducing the fanout. We present several examples of rating functions and rule sequences. For evaluation of methods solving the SSC problem we use a complete, mature and commercially comparable Open Source RTS game called Wargus. We will measure the performance of each variation of the game tree-based methods. This is done by comparing the performance of the variations with the performance of the built-in module in Wargus. The experiments performed in this thesis show that there is room for improvement in the way SSC situations are handled in Wargus. With appropriate rule sequences and rating methods the game tree-based methods perform better than the rule-based systems currently in use. i Resumé Dette speciale præsenterer og analyserer small scale combat (SSC) situationer i realtids strategispil (RTS). Et RTS spil er en krigssimulator hvor forskellige fraktioner kæmper i en virtuel verden. SSC situationer opstår når soldater, kaldet enheder, fra forskellige fraktioner kæmper i denne virtuelle verden. I kommercielle RTS spil håndteres SSC situationer ved at anvende regler til at styre hver enhed, såsom “angrib den nærmeste fjende”. Resultatet er ikke optimalt, som flere eksempler vil vise. Derfor vil vi undersøge andre metoder til at håndtere SSC situationer. Computerspilsbranchen bruger navnet Game AI til at betegne moduler, som håndterer computerstyrede karakterer. I dette speciale vil vi præsentere hvad dette begreb dækker over mht. metoder og algoritmer. Denne undersøgelse inkluderer et overblik over udvalgte computerspilsgenrer, en introduktion til hvad computerspilskvalitet er, samt et overblik over hvordan Game AI bliver brugt indenfor forskellige genrer. Vi vil introducere begrebet konsistente karakterer og på dette grundlag vil vi diskutere, hvorledes konsistens påvirker et computerspils kvalitet. Vi vil argumentere for, at hvis RTS enheder skal betragtes som konsistente så skal de opføre sig på en optimal måde med hensyn til den virtuelle verdens regler. I dette speciale præsenterer vi en metode til håndtering SSC situationer. Metoden baseres på et regel-induceret tidsstemplet spiltræ. Grundet mængden af information i spiltræsknuderne vil vi kigge nærmere på en maskinindlæringstilgang til at afgøre knudernes værdi. Vi reducerer antallet af børn til hver knude ved at anvende sekvenser af regler på disse. Vi præsenterer adskillige eksempler på værdifunktioner og regelsekvenser. Til at evaluere metoder der håndterer SSC situationer, bruger vi et komplet og kommercielt sammenligneligt Open Source RTS spil, kaldet Wargus. Eksperimenter der viser spiltræsvarianternes ydeevne er udført ved at evaluere hver spiltræsvariation på flere konstruerede SSC situationer mod det indbyggede modul i Wargus. De målte resultater viser at den måde Wargus håndterer SSC situationer på kan forbedres. Med egnede regelsekvenser og værdifunktioner kan en spiltræsbaseret metode yde bedre end de regelbaserede systemer der bliver brugt i RTS computerspil idag. ii Preface About the authors The authors of this masters thesis are Ronni Laursen and Daniel Nielsen, both masters students of computer science. Ronni holds a bachelor in multimedia and computer science; while Daniel has a bachelor in mathematics and computer science. Both authors gained an interest in computers and computer games in the late 1980’s and that interest has increased ever since. The natural curiosity of what comes next in the market of computer games is a big part of being in the computer game “sphere”. Both authors take great pleasure in following the media outlets dealing with computer games, to glean the latest news about new releases and technologies. Daniel’s primary interests in computer games are real time strategy games, role playing games and the ever evolving market for console games. Ronni’s interests are primarily first person shooters and role playing games. Acknowledgements A big thank you from both authors goes to Ole Caprani for supervising us when everybody else had left. Thiemo Krink for countless fruitful discussions. Brian Mayoh for several references and discussions. Daniel also wishes to thank his family, Christine and Marie, for support and for putting up with the long hours that went into the making of this thesis. Ronni wishes to thank his family for support and love. A special appreciation goes to all of my close friends for your support throughout the years. Love you all. iii