Another implementation of the game of Kalah. I implemented Wikipedia's rules exactly, as they were also what I was accustomed to playing.
$ git clone https://github.com/bediger4000/kalah.git $GOPATH/src/kalah
$ cd $GOPATH/src/kalah
$ go build kalah.go
$ ./kalah
OR
$ ./kalah -M
You don't have to install it anywhere - it runs in place. It has no configuration file(s).
The "-M" for Monte Carlo Tree Search is probably a more exciting opponent. The Alpha/Beta version just seems cold-blooded and relentless.
Players declare their moves with a number. The usual 6 pits per player board is represented like this:
computer
5 4 3 2 1 0
X Y
0 1 2 3 4 5
human
Computer's pot or store is X, human's is Y. Players move by indicating which of their pits (0 through 5) they want to move. The program empties the chosen pit, then distribute its contents (stones or seeds), one per pit, traveling counterclockwise. The program drops a stone in a player's own pot, but not their opponents. Players get a bonus move if they drop the final stone in their hand into their own pot (X for computer or Y for human, above). A bonus move can result in a bonus move.
This game does do captures. If a player drops their final stone in an empty pit (0 through 5) on their own side of the board, that last stone, and any stones in the pit opposite that empty pit get moved to the appropriate store.
kalah the program displays the current game board,
then asks the human to input a move, which is a single-digit
number, 0 through 5.
Command line flags:
-C Computer takes first move
-M Use MCTS instead of alpha/beta minimax
-P Do CPU profiling
-R Reverse printed board, top-to-bottom
-U float
UCTK factor, MCTS only (default 1.414)
-d int
lookahead depth for Alpha/Beta, moves for each side (default 6)
-i int
Number of iterations for MCTS (default 200000)
-n int
number of stones per pit (default 4)
"MCTS" means Monte Carlo Tree Search. It defaults to deciding what move to make by using Alpha/Beta minimaxing.
Reverse printed board makes it easier to open two terminals side-by-side and play instances of the game against each other. Use "-R" on one of the two instances so the programs print boards that look the same.
There's nothing magic about minimax using 6 move look ahead, or Monte Carlo Tree Search using 200,000 iterations. I found them empirically. Both move choice algorithms can usually beat me if they move first, or if I make a mistake, and neither algorithm takes too long to decide with those values.
I don't see that varying the UCTK parameter makes any difference. You can't use 0.0 as a value, however.
I used the Wikipedia article on Alpha/Beta minimaxing. I should have implemented one level of threading. Static value calculated as difference of player's pots or stores.
I used the Wikipedia article on Monte Carlo Tree Search for the MCTS algorithm. The difficulty lies in the untried moves. Dropping a final stone in a player's store means that player makes the next move, so calculation of untried moves depends on the "next player".
This MCTS does a lightweight playout. Once the Expansion part of the algorithm is complete, the code just does random legal moves until someone wins.
This variant has a bonus move. Players get an extra move if they drop the last stone in their hand into their own pot/store. The Alpha/Beta minimaxing code takes care of this in recursion by keeping the "next" player is the same as the current player, and not incremented ply count. It doesn't reach its move horizon while a player is in the middle of making a multi-move sweep. This does lead to unexpected increases in move calculation time during mid-game, when a lot of bonus moves occur.
I wrote another program to try one algorithm against another.
$ git clone https://github.com/bediger4000/kalah.git $GOPATH/src/kalah
$ cd $GOPATH/src/kalah
$ go build playoff.go
$ ./playoff
4 4 4 4 4 4
0 0
4 4 4 4 4 4
> Hit return at ever > prompt to see the next move.
Player 1 is at the top, player 2 has the bottom row of pits.
Although Alpha-beta minimaxing can handily beat a human at a depth of 6 moves (12 plies), MCTS+UCB1 can beat A/B minimaxing looking ahead to a depth of 7 moves, even if MCTS goes second.
I must have something wrong with the static valuation function.