An attempt to train a NN for the game Briscola Chiamata using rllib (still very preliminary).
Current status:
- The game differs from the actual one in 3 main aspects:
- the bidding phase stops when the bidder has offered "2", the lowest ranked card in the deck. It is not possible to bid further, on the number of points.
- As a consequence, the rewards are not differentiated based on the total game points achieved by a team. "Cappotto" (when one of the teams gets all the available 120 points) is not implemented either.
- The trump suit is chosen by the caller right after the bidding phase, and not after the first hand has been played
- DebugGame.py is able to run a game among 5 RandomAgents
Next immediate goals:
- Train a NN with these rules and check if it is able to systematically beat a RandomAgent on a sufficiently
big number of games (1000?)
- to achieve it, it is necessary to sensibly choose the observation space. A reference can be the work done for a similar German game, Schakopf
- Implement a human player and a game against a mix of NN and Random agents