Partial Computing is a research and practice lab for games of incomplete information.
Elections, negotiations, markets and alliances are played by agents who each see only part of the board. We study how these games unfold, how technology is changing the way they are played, and where it could help resolve them.
We work across game theory, design, neuroscience and philosophy, because understanding how agents interact means looking at what they know, what they perceive and what they value.
Practice
Bounty
Artificial agent ecology
Ongoing
An online social game for 20 players, humans and AIs mixed without revealing who is who. Over eight rounds they explore a map, secretly steal loot, negotiate, lie, accuse each other and vote to punish others. Whoever collects the most points wins.
The real aim is experimental: to compare how humans and AIs behave, measuring inequality, movement, alliances, votes and strategies, and to find out whether they can actually be told apart.
Minimax Solver
Computational game theory
Ongoing
An experimental research solver for imperfect-information games such as no-limit Texas hold'em. It studies the metrics behind the GTO standard, like exploitability, and searches for alternative algorithms that reach the same quality with less compute.
A spot is defined in YAML (ranges, stacks, bet sizes, rake) and the solver builds the game tree, solves it with vectorised CFR over all 1,326 hands on CPU or GPU, and tracks convergence. It covers preflop and postflop, heads-up and multiway, cash games and ICM. To save compute it tests value networks at the leaves of the tree and memory compression, with benchmarks comparing accuracy, speed and resources.