Sei sulla pagina 1di 1

Report 11: Ngo, Thuan Van November, 26th, 2015

Question 11.
(a) How can you design screen type games? For example, chess, shogi, go, checkers, etc.

Kriegspiel (screen chess), [1] in this game each player can see their own pieces, but be blind to those
of their opponent. For this reason, it is necessary to have a third person (or computer) act as a
referee, with full information about the progress of the game. When it is a player's turn he or she
will attempt a move, which the referee will declare to be 'legal' or 'illegal'. If the move is illegal, the
player tries again; if it is legal, that move stands. Each player is given information about checks and
captures. They may also ask the referee if there are any legal captures with a pawn. Basing on this
information, we could define any screen-type game that need to possess three factors: a player is
unable to see the others moves, the game needs a referee to judge each players movement, the
referee could provide some other essential information to players.
(b) What would be your strategy to play such screen type games? For example, in the stage of
opening, middle game and endgame.

In my opinion, Monte Carlo tree search (MCTS) [2] could still represent its effectiveness in solving
such kinds of problems. By considering four steps in MCTS-based technique [1]:
-

Selection: The algorithm selects a leaf node from the tree based on the number of visits and
their average value. Selection is a problem reminding of the n-armed bandit, where the agent
needs to keep a balance between exploration and exploitation.
Expansion: The algorithm optionally adds new nodes to the tree.
Simulation: The algorithm somehow simulates the rest of the game one or more times, and
returns the value of the final state (or their average, if simulated multiple times).
Backpropagation: The value is propagated to the nodes ancestors up to the root, and new
average values are computed for these nodes.

We could apply MCTS as follows:


-

For Selection: using a maximizing, or minimizing, or optimal strategy to visited nodes in the
goal to reach the leaf.
For expansion: the expansion of a new node depends upon a specific traversal, however it
could be considered when we really need the extra information.
For simulation: Random is a possible tactic.
For backpropagation [3]: Any possible solution is acceptable.

Although it may be only a theoretical idea, we hope it could help in any potential way.

References:
1. Monte Carlo tree search in Kriegspiel - Paolo Ciancarini, Gian Piero Favini
2. Monte Carlo tree search from Wikipedia.
3. Backpropagation from Wikipedia.

Potrebbero piacerti anche