Speakers
Description
This paper aims to improve existing solutions of machine-learned bots for the Blood Bowl game. Blood Bowl is a stochastic, fully-observable, turn-based, two-player board game. Players, referred to as coaches, lead an 11-player team, where each individual has statistics and abilities. A wide grid-based board, various moves, possible interruption of the game sequence, and randomness create a complex and challenging environment. For such, most of the state-of-the-art solutions are scripted bots.
To accomplish our goal, we used Open Source and predefined code. Our solution was based on the first successful, deep-learning-based bot for Blood Bowl – MimicBot, the BotBowl III winner. It used a hybrid Reinforced Learning (RL) and Imitation Learning model. We used Behavioral Cloning and A2C models to train it, which offer a balance between simplicity and efficiency. We evaluated MimicBot, testing different settings and observing the reactions. Based on that, adding modifications to increase the bot's capabilities was possible.
