Artificial intelligence in play mode: How AI masters the card game «Cyber Shield»
«Cyber Shield» is the new serious game of the Swiss Armed Forces for raising awareness and building basic understanding in the areas of cyber as well as electromagnetic security and defence. Up to now, this game has only been developed and analysed using time-consuming human versus human games. In order to improve this process, artificial intelligence which can take over the role of one or more players is currently being developed at armasuisse Science and Technology (S+T).
Matthias Sommer, Operations Research and Systems Analysis at armasuisse Science and Technology, Stefan Lehmann, Cyber Protection in the project Cyber Command
Introduction to the game «Cyber Shield»
Cyber Shield is a strategic two-player game, i.e. 1 against 1. The aim is to securely increase the scope of your own networks. The more robust the design of a player’s networks, the more points their networks are worth. Numerous treacherous cyber threats can be used to weaken the opponent’s networks and to reduce the score. The first player to achieve a score of 10 or more points at any time wins the game immediately. A player who cannot take a card or whose total score is -10 or less at any given time loses the game immediately. The players alternate as active and passive players and take new cards from their own pile of cards as soon as it is their turn. The cards are then played from the players’ hands and placed in the middle of the table in their own network area. Each round, ever more complex networks are created and the threat scenario increases continuously: an unrelenting race.
Problem
One difficulty in the development of strategic games is to keep the game balanced while at the same time exciting. Thus, for example, none of the cards may be so dominant that their use would lead to unavoidable victory. In complex games, such as «Cyber Shield» of the Swiss Armed Forces, it is not clear right from the beginning whether the game is well-balanced, for example, whether no card is too dominant or an individual strategy always leads to victory. The combinations are thus very numerous even with a relatively low number of cards. To find this out, games played by humans have been analysed up to now. However, as each game is time-consuming and the results are only available in analogue form, the author of the game is not able to make a final assessment of the many special cases which only occur very rarely.
Thousands of games played in a very short time thanks to digitisation
It is therefore more efficient to digitise the game and then develop software agents which can play the game with strategic foresight. A method that can do just this is what is known as reinforcement learning. This method from the area of artificial intelligence learns the winning strategy of a game through «trial and error» and can then apply this in games against other software agents or even against human players.
In the case of «Cyber Shield» one employee from armasuisse Science and Technology and one from Defence implemented the process described above in cooperation with an external research partner. Initial tests show that the developed «intelligent agents» can compete with the average human player. It is thus now possible that thousands of games could be played in a very short time and at the same time, each played card could be recorded.
Expected findings
As the project was only launched in autumn 2023, no reliable findings are available yet. What is already becoming apparent and to a certain extent the main result is that the game cannot be «broken». This means that there is no all-dominating winning strategy. Further possible statements or characteristics which need to be considered could be, for example:
- There is no statistical correlation between the starting player and the chance of winning: both players win the same number of times.
- The average number of moves up to the win is 35, although there are games with only 6 or up to 60 moves.
- Development of game scores (points), see Figure 1.
- Frequency of the cards played per game, in other words, how often each card has been played.
- Average points earned for each individual card, in other words, how a played card affects the score of both players.
In addition, the strategies of the agents can be parametrised. For example, the level of aggression with which an agent should proceed can be specified. The connection between aggression and «winning chance» can be derived from this.
These findings help the further development of the «Cyber Shield» game. More complex cyber aspects can thus be integrated, which means that the troops’ awareness of these cyber aspects is raised. Ultimately, this also promotes basic understanding in the area of cyber security.
Cyber Shield – The serious game on cyber security (admin.ch) (in german)