I am very happy to share that my Cheqqers AI won the Cheqqers tournament at the 1st European Tournament in Quantum Games.

The tournament took place on 20–21 June 2026 in Maastricht, the Netherlands, and brought together researchers, students, and enthusiasts working at the intersection of quantum computing, artificial intelligence, and games. Organized in connection with the International Computer Games Association, the event focused on quantum games such as TiqTaqToe and Quantum Checkers, also known as Cheqqers.

For me, this was a particularly exciting setting: games have always been an important testing ground for artificial intelligence. Chess, Go, checkers, and many other games have helped shape the development of search algorithms, reinforcement learning, and decision-making systems. Quantum games now open up a new version of this playground, where classical game-playing ideas meet concepts such as superposition, uncertainty, and probabilistic state representations.

What is Cheqqers?

Cheqqers is a quantum version of checkers. At first glance, it resembles the familiar board game: pieces move diagonally, captures matter, and promotion still plays an important strategic role. However, the game introduces quantum-inspired mechanics that make the state space much richer.

Instead of every piece simply being present or absent on a square, positions can involve probabilities. A piece may be partially present in different locations, and moves can create quantum splits. This means that an AI player cannot simply treat the board as a classical checkers position. It has to reason about uncertain occupancy, partial captures, and the value of pieces that may only exist with a certain probability.

That makes Cheqqers a fascinating benchmark for AI. It is still structured enough to allow strategic search, but complex enough to force new representations and heuristics.

My approach

For the tournament, I developed a game-playing AI for Quantum Cheqqers on a 6×6 board with quantumness level 2. At first, I tried a self-play approach with a transformer-based estimator and a precomputed set of strong first and last moves. However, after only one night of training, it was not competitive enough, reaching about a 75 percent win rate against random.

With only two hours left before the tournament, I changed direction and coded the final agent: a quantum-aware Monte Carlo Tree Search. This quickly achieved a 100 percent win rate against random and became the basis of the tournament-winning Cheqqers AI.

The central idea is simple: rather than trying to exhaustively evaluate every possible continuation, the AI repeatedly simulates possible futures from the current position. Over many playouts, it learns which moves tend to lead to better outcomes.

The search follows the usual MCTS structure:

1. select promising moves in the current search tree
2. expand the tree with new candidate moves
3. run guided rollouts to estimate the outcome
4. backpropagate the result through the tree

For move selection, I used a PUCT-style formula that balances exploitation and exploration. This allows the AI to spend more time on moves that already look strong, while still investigating alternatives that may turn out to be better.

After the simulations, the AI chooses the root move that was visited most often, with ties resolved by the average value of the move. This makes the final decision robust rather than overly dependent on a single lucky simulation.

Making the search quantum-aware

The most important part of the system is not only the search itself, but the way the AI evaluates quantum positions.

In classical checkers, a piece is either on a square or it is not. In Cheqqers, that assumption no longer works. My AI therefore evaluates material in a probability-weighted way. A piece that is present with 50 percent probability contributes only part of its value. Captures are also evaluated by the amount of probability mass they remove from the opponent.

The heuristic considers several factors, including material, advancement, promotion potential, central control, friendly support, and tactical danger. Quantum split moves are treated as real strategic options rather than ignored or simplified away.

This is what made the agent effective: the MCTS did not search over a classical approximation of the game. It searched over a representation that respected the probabilistic structure of Cheqqers.