Why computer scientists have been fascinated by mancala since the 1960s
In the 1960s, researchers at MIT and Stanford used Kalah to test the minimax algorithm with alpha-beta pruning. A mancala board state could be represented with simple integer arrays, the branching factor was smaller than chess, and games were deterministic. The same techniques, refined over decades, now form the backbone of every chess engine, Go AI, and the bot you play against on Toguz Arena.
Kalah (6,6) was solved in 2002. Here is what "solved" actually means
In 2002, a team led by Geoffrey Irving, Jeroen Donkers, and Jos Uiterwijk published the solution to (6,6)-Kalah. Their proof showed that with perfect play from both sides, the first player always wins by exactly 8 points. The solution required analyzing roughly 10^9 positions using retrograde analysis: working backward from known endgame states, labeling each as a win, loss, or draw for the player to move.
But "solved" has a specific mathematical meaning. It does not mean every human Kalah game is predetermined. A practical player must count ahead, evaluate captures, and decide between bonus-turn chains — all under time pressure and with imperfect reasoning. As one Reddit player put it: "Most people aren't making all the optimal moves. Otherwise no one would ever play with them."
Oware and the 2002 Awari Computer Olympiad
Oware attracted AI researchers because of its manageable branching factor (no bonus turns) and the feeding rule's unique constraint. The 2002 Awari Computer Olympiad was won by a program using retrograde analysis — pre-computing perfect play for all positions with 35 or fewer seeds on the board, then using minimax for earlier phases. This hybrid approach — perfect endgame + heuristic opening — is the template for strong mancala AI today.
Togyz Kumalak: why it has never been solved
The 9x2 board, 162 stones, and the tuzdyk rule create a fundamentally different challenge. Once a tuzdyk is established, the board state cannot return to its pre-tuzdyk configuration. This irreversibility means standard retrograde analysis is not applicable. The search space is estimated at roughly 10^12 positions — 1,000 times larger than Kalah. The 2025 World Championship featured a position that three separate engines could not agree on, with evaluations spanning a 15-point swing.
Neural networks enter the ancient game
The first AI-powered Togyzkumalak competition was held in Astana in March 2026. Several entrants used neural network-based evaluation functions trained on games from the World Togyzkumalak Federation archives. Asel Dalieva, the 2024 Women's World Champion, trains with AI analysis tools — reviewing tournament games with an engine, identifying where her evaluation diverged from the machine's, and incorporating the feedback. This human-AI collaboration is the future of mancala training.
Frequently Asked Questions
Can a computer always beat a human at mancala?
At Kalah (6,6), yes — the game is solved, and a perfect computer always wins as first player. At Togyz Kumalak, no — the game is unsolved, and top human players with neural analysis tools are competitive with the strongest engines.
Why was mancala used for early AI research?
Perfect information, deterministic sowing, manageable state space, and clear win conditions made Kalah an ideal testbed for minimax algorithms that later powered chess engines.
