Dynamic di culty adjustment (DDA) in video games context allows a game to be automatically adjusted and customized based on the player's performance. Our goal is to build a DDA architecture able to learn from behavioral data and return appropriate feedback to maximize the player's performance and enjoyment. For our experiments, we use Tetris Analytics video game, which is like a traditional Tetris from the player's point of view. However, it allows us to extract data from the games such as the movements made by the player, the state of the board, next piece, etc. With this data, we predict the player's pro le and personalize the game.
A majority of the game content is static and created during the development
phase of video game production. Game designers attempt to make game content
that is challenging but not overly di cult for the average player in the target
audience. For this reason, feedback from players is essential to help game designers
to distinguish from an exceedingly di cult task to a great challenge [
Dynamic Di culty Adjustment is a new subject compared to other application
areas in the eld of video games. It can be used to automatically alter the game's
di culty to meet the player's expectations [
Our research focuses on using Theory of Flow, DDA, CBR and time series in video game context to create a DDA architecture that improves the experience of the player by adapting the game to the level where it is challenging but not overwhelming. 2.2
Speci c Objectives { Review DDA state of the art in video game context. { Evaluate novel techniques for the categorization of behavioral data. { Use time series to identify the evolution of a player in a speci c time range. { Use categorization techniques for player pro le prediction. { Review The Theory of Flow and its application in video game context. { Implementation of DDA in real game. { Execute experiments to analyze DDA efect in players' experience. { Create DDA architecture that allows abstraction of the fundamental processes for game adaptation. 3
The planning of our research divides into three stages. The rst stage focuses on the review of the state of the art and application of knowledge acquired in a simple video game called Tetris Analytics. This stage takes place in the rst two years, and it is complete. The highlighted activities of this phase are the following: { Review of the state of the art of machine learning techniques. { Review of techniques applied in video games for players categorization and prediction of their skill level. { Perform experiments of player classi cation and dynamic di culty adjustment based on user interaction in TetrisAnalytics using the techniques reviewed previously. { Compare performance of evaluated techniques. { Publication of ndings.
The second stage focuses on the establishment of a methodology for players categorization and dynamic di culty adjustment based on user behavior in video games. This stage is partially completed and must nish next year. The main activities are: { Establish general processes used in Tetris Analytics that can use in other games. { Establishment of a methodology for a player's categorization and dynamic di culty adjustment in video games. { Publication of ndings.
The third stage focuses on the application of our ndings in a more sophisticated video game. This stage should nish within two years, and we have not started yet. The main activities of this stage are: { Conducting experiments with a more complex game for players categorization and dynamic di culty adjustment based on user interaction. { Veri cation of methodology. { Publication of ndings. { Writing the thesis.
{ Preparation and presentation of the thesis. 4 4.1
We used Tetris Analytics, a version of Tetris implemented in Java by a member of our research group. Our Tetris game looks like the original one for the player point of view but internally provides extra functionality to extract, store, and reproduce game traces. Tetris is a widely known game, where the di culty is always rising, so it is easier to evaluate the stress level of a player, and it is a simple game to adjust. Therefore we can extract behavioral data, select the most signi cant features to characterize the playing style of each player, determine their skill level, and dynamically adjust the di culty of the game.
Each time a new piece appears on the top of the board, players have to make two di erent decisions. The rst one, the tactical decision is to select where to settle the piece. The second decision involves all the moves (translations and rotations) required to lead the piece to that nal location. We only consider the tactical decisions to de ne the skill level of the player, but we are aware that we could also extract valuable information from the concrete moves. 4.2
We used clustering to analyze data extracted from TetrisAnalytics. My work
on this resulted in a publication at The Florida Arti cial Intelligence Research
Society [
{ Flow level: a score describing how good the game was for the player. Note that the game computes the last two features once it nishes. After a game ends, we asked the player on how was their experience during the game. The ow level was a score given by the player on a Likert scale of 5 values.
To calculate player pro le during gameplay, we use a query case Q describing the evolution of the variables during the last ten tactical decision and use a K-Nearest Neighbor classi er (k-NN) algorithm to get similar cases from our general case base. The similarity between the two cases is the linear combination of the similarities between their time series. We use a simple similarity measure based on the Euclidean distance to compare time series. The similarity function only considers cases that have the same time range in previous games. Each time series has a weight or importance. The most signi cant time series is the score (70%), followed by the number of holes in the game board (25%) and nally the max pieces' height (5%).
A majority vote decides the current player's pro le among the k most similar cases. To avoid extreme pro le changes, we use an inertia function that eases the transition from newbie to expert and vice versa. Every time the game predicts the player's pro le, the inertia function is applied, and it compares the current player pro le with the new one. 4.3
To select an excellent piece for the player, we compute all possible ways to place each type of piece and use a heuristic to decide how good those likely future boards are. The heuristic tries to maximize the score (i.e., complete newlines) and minimize the number of holes and height of the board. Then we assign to each piece the rating of the best game board that can be reached placing the piece optimally. Finally, we randomly choose one of the three best options and use it as the next game piece. Now, we only worry about making the game more relaxed, but it is in our plan also to make it harder for more experienced players. Now, we are working on the second stage of the research plan, trying to extract the main processes from dynamic di culty adjustment in Tetris Analytics and create a DDA system architecture. Ideally, from this architecture, we will create an abstraction of the processes involved in dynamically adjusting a game.
Each user has a di erent way of playing, performance level and thoughts
in what would be a game that he would enjoy to the fullest. The idea is to
create a stress level that maximize the user experience during gameplay. For
this purpose, we create a user experience ow that takes advantage of dynamic
di culty adjustment bene ts. In Tetris Analytics we change the game by giving
the best next piece that helps the player make lines [