 |
| Working on a decent 1 vs 1 computer-player that isn't reinforced learning. Even though DeepMind/DeepNash plays Stratego well, which is a game with an incomplete information environment, non-deterministic, and zero-sum, I do not have anywhere close to those resources to train a model on a 40k tile map with a large variety of units with complex/nuanced interactions on multiple layers. I've had to rely on behavior/trigger scripts, calculating utility, influence maps, threat matrices, meta-data attached to maps (choke points, connection networks, strategic terrain overlays etc.), tactical playbooks, operational templates, objective priorities, build orders/tech paths, resource management profiles, game state/game tree search, etc. etc. etc. to try to best simulate a good opponent. This requires datasets much larger than the actual game units/tile stats used to generate various data structures for game rules. |
