Game AI & experiment design2026King's College London
Cinquillo 2.0
A configurable card-game engine with MCTS and RL agents, used to design a measurably fairer game.
33 variants
A research project that treats game design as an optimisation problem. A fully parameterised Cinquillo engine exposes 33 named rule variants; Random, Heuristic, MCTS and reinforcement-learning agents play them at scale; and a 16-experiment pipeline measures how much of each outcome is skill and how much is seat order and deal luck. The output is not a bot — it is evidence for which rules make the game fair.
ContributionEngine, agents, experiment harness and analysis.
33Rule variants
16Simulation experiments
5,000Games per experiment
209RL state features
42Q-network actions
100KReplay buffer
260Tests passing
How it works
- Engine models the 40-card Spanish deck (Oros, Copas, Espadas, Bastos; ranks 1–7 and 10–12) with legality derived from board adjacency.
- Moves are an abstract base class with PlayCard / RollDice / Pass subclasses — variants compose rather than fork.
- 33 named rule variants, including a fully parameterised dice mechanic with four beneficial and several adverse effects.
- Agents span Random, Heuristic, Monte Carlo Tree Search and a reinforcement-learning agent.
- RL agent uses a 209-feature state encoding, a 42-action Q-network and a 100,000-transition replay buffer.
- 16 simulation experiments, 5,000 games each by default.
- README reports 260 passing tests.