Poker Solver vs Poker Bot: The Difference

People often use the words interchangeably, but a poker solver and a poker bot are different tools. The short version for anyone comparing a poker solver vs a bot: a solver computes a strategy for a defined spot, usually offline, so a human can study it. A bot plays the whole game, in real time, against opponents. This article explains the difference, where modern poker AI sits between the two, and what "GTO" means once there are more than two players.

What a poker solver does

A gto solver is software that computes an approximate game theory optimal strategy for a spot you describe. You give it a starting situation: the ranges both players hold, the board, the pot, the stacks and the bet sizes you allow. It then runs an algorithm, typically a form of CFR (see CFR explained), until the strategy is close to an equilibrium. The output is a table of frequencies: with this hand, bet 33% of the time, check 67%.

Solvers are mostly used by humans to study. They are good for questions such as how often to continuation bet on a dry flop. Their limits come from how they are used:

• They solve one spot at a time. You choose the inputs, and the answer is only as good as the ranges you assume.
• They are mostly used for two players (heads-up pots), because the game tree explodes with more.
• They do not play. Someone has to translate the output into decisions at the table, and nothing in a solver watches the opponent.

What a poker bot does

A poker bot closes the loop. It receives the state of a real hand, with its own cards, the board, the pot and the action so far, and must return fold, check, call or raise. It handles every street and every seat. It never gets to choose the inputs, and it has to answer within a time limit. In the Poker Sharingan arena the limit is 5 seconds per decision, and a fast bot averages far less.

A bot can be simple. A rules-based poker bot is a few dozen lines of Python. It can also be built on solver technology: the bot's strategy comes from the same family of algorithms a solver uses, but it is computed for the full game and stored for use during play.

Where blueprint plus search sits

Modern poker AI such as Libratus, Pluribus and SHARINGAN-2 sits between the two. It uses solver-style computation in two places.

• Blueprint. Before any game, the program plays against itself for a very large number of iterations and stores a strategy for the entire game. This is the offline, solver-like part, but it covers every spot at once, using abstraction to keep the game manageable.
• Real-time search. When a hand reaches a decision, the bot can re-solve the situation in front of it, starting from the blueprint and the actual betting so far. This is the solver-like part done live, on a smaller problem.

Pluribus used exactly this design with depth-limited search, as described by Brown and Sandholm (2019). SHARINGAN-2 follows it, with search on the turn and river and a time budget that falls back to the blueprint. In short, a solver gives you an answer for a spot, a blueprint gives you an answer for every spot at lower resolution, and blueprint plus search gives you a bot that is detailed where it matters. The history is in Poker AI history.

What GTO means with more than two players

"GTO" or game theory optimal poker usually means playing a Nash equilibrium strategy: one that cannot be beaten by any single opponent deviating. In heads-up poker that is a strong promise, because the game is two-player zero-sum. An equilibrium strategy there will not lose in expectation, whatever the opponent does.

With three or more players the promise weakens. A Nash equilibrium still exists, but your equilibrium strategy is only safe if everyone else plays their part of the same equilibrium. Other players can pick different equilibria, or none, and the result can cost you. There is also no proof that CFR converges to an equilibrium in multiplayer games. So "GTO" in a six-player game is better read as "a strategy that is hard to exploit in practice" than as a guarantee. Pluribus showed that CFR-style self-play still produces a strong bot empirically. The details are in Six-player no-limit hold'em AI.

Exploitative vs equilibrium play

An equilibrium strategy ignores who it is playing. That makes it safe, but it also leaves money on the table against a weak opponent. An exploitative strategy does the opposite: it observes how opponents deviate from equilibrium, such as folding too much to raises, and shifts to punish that. It can win more against that opponent, but it opens holes that a smart opponent can use in return.

• Equilibrium bots are stable and hard to exploit. Their weakness is that they do not adapt.
• Exploitative bots adapt and can win more from weak fields. Their weakness is that they can be exploited back, and their estimates of opponents can be wrong.
• Hybrids start from an equilibrium baseline and deviate only when the evidence is strong.

Both kinds can be tested in the same way. Put the bot in the arena, play duplicate deals, and read the luck-corrected score with its confidence interval. If you want to try an approach, build a poker bot, submit it with sharingan submit mybot.py, and compare it on the leaderboard. You can also play SHARINGAN-2 to see an equilibrium-style bot in action. This is for research and bot arenas. Real-money poker sites ban bots.

FAQ

What is the difference between a poker solver and a poker bot?

A solver computes a strategy for a specific spot you define, offline, mostly for humans to study. A bot plays whole hands against opponents in real time and returns one action per decision.

Is a poker solver the same as GTO?

A solver approximates a game theory optimal (equilibrium) strategy for the spot you give it. Its output is only as accurate as the inputs, such as the ranges and bet sizes you allow, so it approximates GTO rather than being GTO.

Is GTO unbeatable in 6-max?

No. The unbeatable guarantee of equilibrium play applies to two-player zero-sum games. With three or more players, playing an equilibrium strategy does not protect you if the others do not play the same one.

Do poker AI bots use solvers?

They use the same algorithms. Bots like Pluribus compute a full-game blueprint offline with Monte Carlo CFR and then run a smaller solver-like search during play.

Should I play exploitatively or with an equilibrium strategy?

Equilibrium play is safer and hard to exploit. Exploitative play can win more against a weak or predictable opponent but can be exploited back. Many strong approaches start from equilibrium and deviate when the evidence is clear.