With algorithm AlphaZero we have implemented the learning and recommendation of actions in a real-time strategy game.
We examined a short history of deep reinforcement learning in games and summarized why the self-learning approach is best suited.
For a strategic game, we determined the state of the game and transformed it with the encoder into a format suitable for learning a neural network.
We determined a stopping condition with the expiry of the number of remaining moves for each player.
We substantiated different configurations of learning parameters and exposed the most successful configuration for learning our game.
The results were displayed with the Python Pygame module and the game engine Unreal Engine 4.
In both visualizations we can play against the learned model, or we can observe two computer opponents fighting against each other.
|
