Abstract
In this work, a bot was developed that solves the Central European game Schnapsen.Schnapsen is a game played by two players, whereby the opponent’s hand is not visible.
It is therefore a game with imperfekt information. In order to be able to solve the game
anyway, an approach was used that combines the search in a subgame with a value function that specifies the values for the leaf nodes. A neural network is used as the value
function. For the search in the depth limited subgame, Counterfactual Regret Minimisation [17] is used. In addition, a Public Belief State (PBS) [5] is used, which contains the
probability distribution over the possible histories of the game in a world state. Since the
solution space of Schnapsen allows numerous possible combinations of game outcomes,
samples of game outcomes are drawn based on the probability distribution resulting from
the PBS, which are then used to build a subgame. In order for the bot to find better
strategies that can be used as training data for the network, it uses self-play.
| Date of Award | 2024 |
|---|---|
| Original language | German (Austria) |
| Supervisor | Stephan Dreiseitl (Supervisor) |