Hierarchical portfolio search in prismata

Many unique challenges are faced when trying to write an AI system for a modern online strategy game. Players can control groups of tens or even hundreds of units, each with their own unique properties and strategies, making for a gigantic number of possible actions to consider at any given state of the game. Even state-of-the-art search algorithms such as Monte Carlo Tree Search (MCTS) are unable to cope with such large action spaces, as they typically require the exploration of all possible actions from a given state in a search tree. In addition to the difficulty of dealing with large state and action spaces, other design features must be considered such as varying difficulty settings, robustness to game changes, and single player replay value.