Hi everyone.
Finished 4th before the final rerun.
My strategy is based on … surprise … MCTS.
My version make use of two Trees, one for each player.
The main difficulty for this is that legal moves depend on other player actions.
I handled that issue in the following way:
- when computing childs of a node, give it all “potential” actions and not only actions that are legal for the current state. I mean by “potential” that it would be legal if opponent actions were differents. For exemple, allow a GROW even if the price is too high, or allow a SEED even if the cell is occupied.
- when making node selection (i.e when choosing a legal move) choose between childs that are really legals for the current state.
The goal was to reduce the branching factor.
To reduce it further I also strongly pruned the SEED actions.
An other important choice (but I am not sure it was a good idea) was to impose an order in actions during a day : first COMPLETE, then GROW for trees of size 2, size 1, size 0 and eventually SEED.
It usually make actions cheaper in this order (GROW price depend on number of trees of higher size).
For the sake of simplicity I also limited the number of action during a day at only one by category (COMPLETE, GROW1, GROW2, etc.).
I noticed that several gamers have used Neural Networks with success. I can’ t wait to read their PM.