Scribd
Upload
590 views16 pages

Pente: David Kron, Matt Renzelmann, Eric Richmond, and Todd Ritland

The document discusses different game tree search algorithms that can be used for the board game Pente, including NegaScout, SSS-2, MTD(f), and Alpha Beta with Transposition Table. It provides brief descriptions of each algorithm and notes that Alpha Beta with Transposition Table performed the best in tests, followed by NegaScout and SSS-2, with MTD(f) and basic Alpha Beta performing worse. Faster search is not enough for solving Pente due to its large branching factor, and other techniques like stronger static board evaluation and move ordering are needed.

Uploaded by

Muthu Vijay Deepak
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
590 views16 pages

Pente: David Kron, Matt Renzelmann, Eric Richmond, and Todd Ritland

The document discusses different game tree search algorithms that can be used for the board game Pente, including NegaScout, SSS-2, MTD(f), and Alpha Beta with Transposition Table. It provides brief descriptions of each algorithm and notes that Alpha Beta with Transposition Table performed the best in tests, followed by NegaScout and SSS-2, with MTD(f) and basic Alpha Beta performing worse. Faster search is not enough for solving Pente due to its large branching factor, and other techniques like stronger static board evaluation and move ordering are needed.

Uploaded by

Muthu Vijay Deepak
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
You are on page 1/ 16

Pente

David Kron, Matt Renzelmann, Eric Richmond, and Todd Ritland

Introduction
Game tree search algorithms in Pente
NegaScout SSS-2 MTD(f) Alpha Beta with Transposition Table

Pente
Tic-Tac-Toe, with 5 in a row, a 13x13 board, and a couple other small twists

NegaScout
Developed by Alexander Reinefeld around 1983 Good: NegaScout dominates AlphaBeta
Nodes searched by NegaScout are searched by AlphaBeta, but the reverse is not necessarily true

Bad: NegaScout re-searches same subtrees


Transposition tables become more important

AlphaBeta vs. NegaScout

AlphaBeta vs. NegaScout

SSS-2
Variant of SSS* Attempts to visit nodes in a best-first fashion Maintains a global solution tree which is generated by the expand function and is refined by the diminish function Nodes visited are a subset of nodes visited by Alpha-Beta In theory, slightly better than Alpha-Beta In practice, comparable with Alpha-Beta

SSS-2
G = {root}; //The global solution tree g = expand(root, +inf); //The "best-so-far" value do { temp = g; g = diminish(root, temp); } until (g >= temp);

= leaf nodes of G after first expand(root, +inf) call

MTD(f)
Improvement to Alpha-Beta Always searches minimal windows
nextCall(Alpha = beta -1 , Beta = beta)

Can only return a bound on minimax value so it needs to converge towards it. Theoretically better than NegaScout Name From MTD(n,f)
Memory-enhanced Test Driver f = guess

//pos = boardState // d = depth // f = guess Int MTDF(node pos, int d, int f){ int score = f; upperBound = +Infinity; lowerBound = -Infinity; while(upperBound>lowerBound) do{ if(score == lowerBound) then beta = score +1; else = AlphaBeta(pos, beta-1, beta,d); if(score < beta) then upperBound = score; else lowerBound = score; return score;

MTD(f)

Alphabeta with Transposition Table


for (all possible moves){ if (board with this move is in hash table){ get alphabeta value from table; } else { alphabeta(this table); store result in hash table; } update best move seen; }

A transposition table is simply a hash table that stores previously calculated board configurations.

10

Alphabeta with Transposition Table


Most difficult part is the hash function. If each board position can be occupied by player 1, occupied by player 2, or empty, we have 3*3*3*...*3 = 3^(number of positions). For a 13 by 13 board, the number of boards is 3^169 4.3 * 10^80 WAY WAY WAY too big.
11

Alphabeta with Transposition Table

So, I made a 3dimensional hash table. 1st dimension is a hash on the middle row, 2nd dimension is a hash on the middle column, 3rd is on the forward diagonal.
12

Alphabeta with Transposition Table


Doing this means that arrays will only be initialized when they are used, reducing a lot of space. This has time draw backs however (big arrays are initialized during crucial game 13 time).

Results
Algorithm Minimax AlphaBeta NegaScout SSS-2 MTD(f) AlphaBeta with TT Number of Wins 5 10 10 6 3 2

Conclusion
Fast tree search is not enough!
A branching factor of 169 on a 13x13 board is problematic Pente actually uses a 19x19 board

Need better SBE, move ordering,

Demo (as time permits)

You might also like