Do you feel good and in the zone? Or maybe you are hot and upset? Irritable and frustrated? Or maybe sad and melancholy? While there are all kinds of games for many varieties of moods, it might be a good idea for a video game to adjust its difficulty based on how the player is feeling. Because continually feeling angry at a game may not be as much fun or as good for you.
Scientists in South Korea, at the Gwangju Institute of Science and Technology, have come up with a rather intriguing method for just such a thing. The researchers have developed a dynamic difficulty model that would adjust according to players’ emotions and adjust accordingly to ensure player satisfaction is maximized. Because who doesn’t like maximum satisfaction?
Game developers have long known about the necessary balance when it comes to game difficulty and player progression, trying to find a sweet spot that is neither too hard nor too easy to ensure that the gaming experience you feel good. While the settings can usually be changed, this often requires the player to adjust the settings manually. Korean scientists propose something much more dynamic.
Their model consists of training Dynamic Difficulty Adjustment (DDA) agents, using machine learning that has collected data from human players, which then adjust the game’s difficulty to maximize one of four different aspects related to a player’s satisfaction: challenge , competition, flow, and valence.
The scientists used a fighting game for their model and to train their DDA agents, as human players played the fighting game against AI opponents, generating data for the agents, and the humans also had to answer a questionnaire about their experience. Using an algorithm called Monte-Carlo tree search, each DDA agent uses real game data and simulated data to adjust and modify the opposing AI’s fighting style in a way that maximizes a specific emotion or “affective state”.
Associate Professor Kyung-Joong Kim, who led the study, said that one advantage of their approach is that the player does not need to be monitored with external sensors to detect their emotions. “Once trained, our model can estimate player states using only game features,” he said.
The study was small, with just 20 volunteers, but the team said DDA agents produced AI that improved the overall experience for players. However, fighting games offer the most direct feedback, so it begs the question of how it could be used for other types of games, but the professor had an answer for this.
“Commercial game companies already have vast amounts of player data. They can exploit this data to model players and solve various problems related to game balance using our approach,” Professor Kim said.
Their paper documenting the model, “Dynamic Difficulty Tuning Agent Diversification by Integrating Player State Models in Monte-Carlo Tree Search,” will be published in Expert Systems With Applications on November 1. But for those interested, it is now available online and can be found here.
