We present a simple, rapid and efficient method for creating a ghost AI, an Artificial Intelligence that can imitate playing styles of players in fighting games. The ghost AI created can perform combination actions and make decision about any movement in a similar fashion to a player it is copying. We scan a player's battle data, and then create situation-action pair cases for the corresponding ghost AI to use in actual battles. A ghost AI can be created and run swiftly, using small amounts of memory, making it suitable for console games. Our method is general enough to be used in most 2D and 3D fighting games. We carried out our experiment on Street Fighter Zero 3, one of the most well crafted fighting games, using AI-TEM testbed engine.
An Artificial Intelligence (AI) that can copy a player's playing style had been used in many games for some time. In many racing games, a semi-transparent car, which mimics the player's controls from the previous race/lap, is usually available for the player to race against himself. That semi-transparent car is known as a ghost car. It is one of the well known ghost AI systems that players recognize. In fighting games there were various attempts at ghost AIs. Virtua Fighter 4 allows players to train computer AIs to fight like them. Such ghosts can then be assigned to fight another player. However, feedback from players was not good at the time the game was released because it was hard to train their ghosts case by case. Also, at that time, means of ghost AI exchange and distribution were limited. Therefore the idea was not popular. But in recent years, ghost AI has again been used, in Tekken5: Dark resurrection. This time many things have been changed. Players do not need to train their ghosts in a training mode, they just play the game normally and the system will mechanically create their ghosts. The created ghosts will be used as computer characters randomly when other players play the game. The ghost owners gain points if their ghosts defeat other players. This method makes fighting games more interesting because there will be many fighting styles for computer controlled opponents. Points gained by winning with their ghosts also motivate players to create more ghosts. Any player can fight with an expert without having to meet or make an online appointment. Furthermore, advancement in network systems allows people to easily exchange their ghosts and form ghost AI communities. (http://www.tekkenzaibatsu.com/forums/ghostlist.php) Despite the fact that the ghost AI system is being acknowledged as the definitive AI for fighting games, the method for ghost AI creation remains undisclosed. In this paper, we propose a method for ghost AI creation using data obtained from game memory. Our method can be used in most fighting games. It also requires very small amounts of memory and therefore is suitable for console games.
The main concept of our ghost AI creation is case based AI construction. We extracted a player character's reaction in various situations from battle log data created while playing, then created situation-action pairs for the ghost of that character. Our experiment was made using SFZ3 training mode with character Ryu versus Ryu. AI-TEM was modified to suit our experiment. The ghost AI creation processes are displayed in figure below. The following subsections describe each component in the process.
Original player playing video (Dark green character)
http://www.youtube.com/watch?v=GOIbze1pXyg
mirror
http://www.siamtube.com/view_video.php?viewkey=4b1869781beb89f7b73b
Zip-WMV download (10M)
http://www.cp.eng.chula.ac.th/~g48wth/gai/MeComboB.rar
Ghost AI created by that player video (Dark green character)
http://www.youtube.com/watch?v=nf6JY0xaUyU
mirror
http://www.siamtube.com/view_video.php?viewkey=48307e6f653a2005c9dd
Zip-WMV download (10M)
http://www.cp.eng.chula.ac.th/~g48wth/gai/GhostMeComboB.rar