On-line Adapting Games using Agent Organizations Joost Westra, Hado van Hasselt, Virginia Dignum, Frank Dignum Abstract— Almost all computer games that are currently created use fixed scenarios or simple fixed rules to define the course of the game, which mostly results in very predictable and inflexible behavior of all the elements in the game. Current research done on dynamic adjustability in games already makes it possible for different elements to adjust to the player. However, these approaches are still using centralized control. The serious games we are investigating are constructed using complex and independent subtasks that influence each other. Using centralized control becomes impractical if the complexity and the number of adaptable elements increase. We suggest a multi-agent approach for adapting serious games to the skill level of the trainee. Using separate agents makes it easier to guarantee the natural progression of each element of the game and thus its believability. The user task is selected based on a combination of the possible situations that can be provided by the agents at that stage of the game. The task selection is thus dependent on the user model, the agent preferences and the storyline of the game. The storyline and other requirements are specified by using an agent organization framework. An update function for the user model according to the performance of the trainee, the difficulty of the task and the amount of influence of each subtask is also given. I. I NTRODUCTION Computer games are increasingly being used for training purposes. Games created with the purpose of training people are called serious games. These training programs need to be suitable for many different people and therefore the games need to be adjustable. Currently this adjustment is either done externally by experts that need to guide the adaptation, or the application has a number of predefined levels. On the one hand, the use of expert guidance to adaptation results in quite effective training. However such experts are rare and expensive, and it is not always feasible to have an expert available for each training session. On the other hand, the use of predefined adaptation levels may lead to less optimal adaptation in the case trainees do not fit well with the expected stereotypes. To optimize learning it is beneficial to (automatically) adjust the game while the trainee is playing without the need of an expert guiding this process. This adaptation during gameplay is called online adaptation. Even though many commercial games do not use any adaptation [1], already some research has been done on adaptation in games. However, most of this research focuses on adaptation of certain simple quantitative elements in the game. For example better aiming by opponents or adding more or a stronger type of opponents. Joost Westra, Hado van Hasselt, Virginia Dignum and Frank Dignum are with Universiteit Utrecht, Padualaan 14, Utrecht, Netherlands (phone: +31-30-253-4432; fax: +31-30-253-4619; email: westra@cs.uu.nl,hado@cs.uu.nl,virginia@cs.uu.nl, dignum@cs.uu.nl). There are three important aspects [2] that have to be con- sidered when performing online adaptation. First, the initial level of the player must be identified. Second, the possible evolutions and regressions in the player’s performance must be tracked as closely and as fast as possible. Third, the behavior of the game must remain believable. In this paper we will mainly concentrate on the third aspect while trying to achieve the first two. Believability has a lot different aspects, one of the main aspects is that the behavior of the characters should be long-term coherent [3]. Serious games usually follow stricter scenarios than com- mercial games, because they have to guarantee some useful learning experiences for the trainee. The ordering of differ- ent tasks for the trainee are thus more important than in commercial games. On the other hand, we want to provide the trainee with significantly different, possibly predesigned, believable tasks that are created to optimize learning (the second important aspect of online adaptation). However, al- lowing unlimited and unorganized adaptations quickly leads to a disturbed storyline and the believability of the game will be diminished. The characters and other adaptable elements have to remain consistent because they are usually active for relatively long periods in serious games. For example, a serious game can provide a training task for a fire commander that needs to make sure the victims in a burning building are saved. This task is influenced by a lot of different characters. The victims could be more or less mo- bile or they could be located in simple or difficult locations. There may be bystanders that could obstruct the medics. The police could autonomously control the bystanders or could only act if ordered by the fire commander. Also the behavior of the medical personnel affects the difficulty of the task. As becomes clear from this example the user task is very dependent on the behavior of all the characters in the scenario. Because the tasks that the user needs to perform follow from the behavior of all the different characters the behavior of the characters needs to adapt to the user skills. This means that the characters behave differently on the same task dependent on the skill level of the user. This gives a lot of different possible variations because for each subtask there are multiple variations and they can be combined in multiple ways. One difficulty in this adaptation is that the performance of these type of subtasks can not be measured separately because all the behaviors influence each other. If all agents are allowed to adapt to the user without any coordination, situations will occur where all agents adapt at the same time and thus create unwanted scenarios for the trainee. For example, the victims become less mobile, the police agent becomes less autonomous and the bystanders 978-1-4244-2974-5/08/$25.00 ©2008 IEEE 243