Student(s): Cas Giepmans, Chris du Toit, Kimberly Stessen, Viktor Cef Inselberg, Zsolt Harsányi
Supervisor(s): Cameron Browne & Dennis Soemers
Semester: 2020-2021
Problem statement and motivation:
An important part of designing a new game, is checking whether it is fun and interesting to play. Normally, you test whether the rules are clear, if there’s a lasting interested in playing the game, if losing players can get back on their feet again, and if the majority of matches played end in either wins or losses; not in draws. When a player notices that a match can’t be won anymore, the player might instead want to try to achieve a draw. But what if someone ONLY wants to achieve draws, i.e. has no interest in winning? Well, if the rules of a game don’t account for such intentions, it might be very easy to force a draw whenever a player wants. This is categorized as “non-competitive behaviour”, which is unwanted in basically all competitive games: it often leads to lengthy and boring games that nobody will end up playing. Games that minimize the chance for a draw to occur are called “decisive”, and it is this decisiveness, amongst other things, that we in our project try to measure and quantify.
In a perfect world, every new game would be tested rigorously for its decisiveness through the playing of a large number of matches, between players of different skill levels. Since this kind of testing isn’t always available (but computational resources usually are), testing with Artificial Intelligent players could be a viable alternative. The Ludii General Game System, is a digital framework that has the potential to facilitate this for any and all classic games (board games, card games etc.). Ludii is created for designing, playing, testing and evaluating all kind of games under a single interface. It even has several already implemented AI players that can be used for testing. Currently, Ludii has implementations of over 350 different games. To read more about Ludii, please visit https://ludii.games/index.php.
Every AI player in Ludii is currently configured to try to win. This means that they can be used for regular testing of games, but as described before other “motives” are needed to fully evaluate a game and its characteristics. These “standard” AI players might be so focused on winning, that they overlook secondary strategies which potentially exploit flaws in the game’s design. These strategies could enable a player to never lose, as an example (but also never win). Our goal during this project is to alter these kind of motives of play for the AI players, such that games can be tested reliably. The possible outcomes to any board game are “win”, “loss” and “draw”. On the basis of these outcomes, we can define five different motives of play for AI players. These five variants are:
- Standard: AI tries to win or at least tries to draw (win > draw > loss)
- Defensive: AI tries to avoid defeat (win = draw > loss)
- Blocking: AI tries to draw (win < draw > loss)
- Aggressive: AI only tries to win (win > draw = loss)
- Suicidal: AI tries to lose (win = draw < loss)
Fig. 1: Standard AI Fig. 2: Blocking AI Fig. 3: Suicidal AI
In figure 1 to 3 above you can see some of these AI’s at work. Our AI’s play the circle-team and the size of the blue circles indicate the confidence of the AI to make that move. Figure 1 shows the standard AI which is clearly trying to win. In contrast to the defensive AI in figure 2. This AI is easily able to win, but rather blocks his opponent to avoid defeat. The last AI in figure 3 shows us suicidal behaviour. This AI is in the position to avoid defeat, but rather choses any other move.
Since this is a research project, we will limit our scope of games to 2-player, (alternating) turn-based, “perfect information” games. Perfect information means that there are no hidden elements (like cards in a player’s hand), and both players can observe every part of the game at any point in time. We also limit ourselves to “deterministic” games, meaning that games containing chance-mechanisms won’t be taken into account. While looking into games that fall outside of these definitions is academically interesting, it is better to start with something achievable.
Research questions/hypotheses:
During this project, our main research question that we try to answer is:
- Can we automatically detect rule sets that allow players to avoid defeat at will?
To answer this question, we deconstruct it into several sub-questions around which the project will be structured:
- Can you detect non-competitive behavior in a series of playouts (game logs)?
- Why and how does a rule set not disallow this behavior?
- How can such a rule set be detected automatically for any board game?
Main outcomes:
The ultimate goal of this project is to produce an AI system that can test various board games for robustness. More specifically, we would like to test that in a 2-player board game, one player cannot force a draw (or avoid defeat) at will.
References:
Cameron Browne. Lecture: Automated game design, 2020; Cameron Browne. Ludii general game system.