Discover how AI solitaire solvers evaluate moves, predict outcomes and improve winning strategies.
An AI solitaire solver is a computational program that takes a solitaire deal as input and attempts to find a winning move sequence — or, if none exists, to confirm that the deal is intrinsically unwinnable. The term "AI" is used broadly here: in practice, the most effective solitaire solvers are not neural networks or machine-learning systems in the modern sense but rather classical search algorithms enhanced with domain-specific heuristics and pruning strategies that eliminate unpromising move sequences early. The distinction matters because the word "AI" implies learning from experience, while classical search solvers work from explicit rules — but both are called AI solvers in the solitaire community, and both provide strategic insights that transfer to human play.
An AI solitaire solver is a computational program that takes a solitaire deal as input and attempts to find a winning move sequence — or, if none exists, to confirm that the deal is intrinsically unwinnable. The term "AI" is used broadly here: in practice, the most effective solitaire solvers are not neural networks or machine-learning systems in the modern sense but rather classical search algorithms enhanced with domain-specific heuristics and pruning strategies that eliminate unpromising move sequences early. The distinction matters because the word "AI" implies learning from experience, while classical search solvers work from explicit rules — but both are called AI solvers in the solitaire community, and both provide strategic insights that transfer to human play.
The core operation of any solitaire solver is the same: it represents the board as a state (the current arrangement of all cards), generates all legal moves from that state (the successor states), evaluates which successors are most promising (using a heuristic function or exhaustive enumeration), and explores successors in priority order until either a win state is reached (all cards on foundations) or all states are confirmed losing (dead ends with no legal moves). The differences between solver architectures lie in how they prioritise which successors to explore, how they detect and prune dead-end branches, and how they handle the hidden information in variants like Klondike where face-down cards create an incomplete information problem that exhaustive enumeration cannot fully resolve.
Understanding how AI solvers work is strategically valuable not because players can implement them in real time — they cannot — but because the solver's architecture reveals why certain human strategic habits are correct: the forced scan sequence approximates the solver's heuristic priority function; undo-based hypothesis testing approximates the solver's backtracking mechanism; circular dependency checking approximates the solver's dead-end detection. Each human habit is a scaled-down, real-time-executable version of a solver component, and understanding the solver component explains why the habit works and when its approximation breaks down.
From an AI solver's perspective, solitaire is a directed graph search problem. Each node in the graph is a distinct board position — a specific arrangement of all cards across tableau, stock, waste, foundations, and free cells (in variants that have them). Each directed edge from node A to node B represents a legal move that transforms position A into position B. The solver's task is to find a path through this graph from the initial node (the shuffled starting position) to any win node (a position where all cards are on foundations in correct order), or to confirm that no such path exists.
The size of this graph varies enormously by variant and determines how challenging the solver's search problem is. For FreeCell, the graph for a single deal has been estimated at billions of distinct nodes in the worst case — but in practice, most winning paths are found by efficient solvers within seconds because the heuristic function can aggressively prune the graph to the few hundred or thousand nodes that are on or near a winning path. For Klondike, the graph is smaller per deal but the hidden information creates a meta-graph: the solver must handle not one graph but the set of all graphs consistent with the possible hidden card arrangements, multiplying the search complexity. For Forty Thieves, the 80-card two-deck state space with restricted build rules creates a graph that is large both in node count and in the proportion of nodes that are dead ends — which is why Forty Thieves has such a high unwinnable rate and why solver analysis of its deals is computationally expensive relative to other variants.
Component 1: State representation. Every solver must define a compact, unambiguous representation of each board position that captures all information relevant to future move generation. A typical FreeCell state representation encodes the card in each of the eight column positions, each of the four free cells, and each of the four foundation tops — approximately 60 values that fully specify the position. A Klondike state representation must also encode the stock and waste pile ordering and the face-down card arrangements, which adds information that may or may not be fully known depending on how many face-down cards have been revealed. The state representation determines how efficiently the solver can store visited positions (to avoid re-exploring already-seen states) and how quickly it can generate successor states from any given position.
Component 2: Move generation. From any state, the solver generates all legal successor states — all positions reachable from the current state by exactly one legal move. The quality of move generation directly affects solver efficiency: a solver that generates all legal moves including moves that are provably suboptimal (such as moving a card to a free cell and then immediately moving it back) wastes time exploring dominated branches. High-quality solvers implement move-generation pruning rules that eliminate dominated moves before exploration — for example, never moving a card from a free cell to a column if the same card could have been moved there directly without involving the free cell, or never placing a card on the foundation if it would prevent a lower-ranked card of the same colour from being placed on the foundation at a later step. These pruning rules are exactly the strategic principles that expert human players apply — they are heuristics that the solver implements as move-generation constraints rather than as explicit strategy choices.
Component 3: Heuristic evaluation. The solver's heuristic function assigns a priority score to each generated successor state, determining which states are explored first. A good heuristic for solitaire assigns high scores to positions with more foundation cards, fewer face-down tableau cards, more empty columns or free cells, and higher degree of suit consolidation in the tableau sequences. The heuristic is the solver's approximation of position quality — and it is directly analogous to the human player's position assessment. The specific features that high-quality solitaire heuristics weight most heavily — foundation advancement, face-down card reduction, empty column preservation — are exactly the features that the strategy cluster's priority frameworks (forced scan sequence, empty column discipline, foundation balance) tell human players to prioritise. The heuristic is not arbitrary: it is empirically calibrated by running the solver on millions of deals and measuring which feature weightings produce the highest proportion of winning paths found within the computation budget.
Component 4: Dead-end detection and pruning. The most computationally expensive part of solving an unwinnable deal is confirming that it is unwinnable — which requires demonstrating that every possible path from the starting position leads to a dead end. Efficient solvers implement dead-end detection heuristics that identify structural blockage patterns (circular dependencies, key card burial configurations) early in the search and prune entire subtrees of the move graph rather than exploring them exhaustively. The circular dependency check — does any pair of cards block each other's movement with no external resolution available? — is the solver's most powerful dead-end detection tool and the one that maps most directly to the human diagnostic habit described in the unwinnable deals guide. Solvers implement this check automatically on every state they visit; human players implement it manually as the first step of the three-pattern structural diagnostic before resigning.
The move-generation pruning rules are the strategic principles. Every pruning rule that a high-quality solver applies during move generation corresponds to a strategic principle that expert human players apply during move selection. The rule "never move a card to a free cell if it can be placed directly on a column" maps to the free cell rationing principle. The rule "never place a card on the foundation if it creates a build base gap that blocks a lower-ranked same-colour card" maps to the foundation balance principle. The rule "always prefer uncovering a face-down card to building a sequence of equal immediate value" maps to the uncovering-first principle. Understanding that these principles are solver pruning rules — rules that eliminate dominated move sequences from the search space — explains why they work: they are not arbitrary conventions but mathematically justified eliminations of move types that have lower expected value than their alternatives across the full deal distribution.
The heuristic function explains what to maximise. A solitaire solver's heuristic function is a linear combination of board features, weighted by their empirical contribution to win probability: foundation advancement (weight highest), face-down card reduction (weight high), empty column count (weight medium-high), free cell occupancy (weight medium-negative). This weighting structure is the objective function that correct solitaire strategy maximises — and it directly explains why the forced scan sequence's priority ordering is correct. Foundation moves score highest on the heuristic because they advance the win condition directly and irreversibly. Uncovering moves score second because they reduce face-down card count, the second-highest weighted feature. Pure tableau builds score third because they increase sequence organisation without directly advancing the heuristic's top two features. Stock draws score last because they consume finite stock resources without directly advancing any heuristic feature — they are necessary but their heuristic cost (reduced stock capacity) exceeds their heuristic benefit (position advancement) in most states where tableau moves are available.
The backtracking mechanism explains the correct use of undo. A solver uses backtracking to explore alternative paths when a promising path reaches a dead end — it returns to the last branching point and tries the next-highest-heuristic successor instead of the branch that failed. Human players using the undo function in online solitaire perform the same operation: when a move sequence leads to an apparent dead end, undoing back to the last meaningful branch point and trying an alternative path is exactly the backtracking step. The key difference is that solvers backtrack systematically — they remember all unexplored branches at every branching point and explore them in priority order — while human players backtrack selectively, using pattern recognition to identify which branches are worth trying rather than exhaustively enumerating all options. Developing better pattern recognition for which alternative branches are worth testing after a dead end is the primary way human undo-based play improves toward solver-level performance.
Solver performance on specific variants reveals which human skills matter most. Solvers solve FreeCell deals fastest (milliseconds), Klondike deals slower (seconds to minutes for difficult positions), Forty Thieves deals slowest for unwinnable confirmation (minutes to hours in extreme cases). This performance gradient reflects the same difficulty dimensions that human players experience: FreeCell's complete information makes the search tractable and correspondingly makes the strategic path calculable; Klondike's hidden information expands the search space and correspondingly makes the correct strategy estimable rather than calculable; Forty Thieves' restricted build rules and large state space make the solver's search expensive and correspondingly make the human player's diagnostic process slower and less reliable. The human skill that matters most in each variant is the skill that the solver's architecture most explicitly implements: complete state evaluation in FreeCell, conditional probability estimation in Klondike, efficient dead-end detection in Forty Thieves.
Believing that a solver's winning path is the optimal human strategy. An AI solver finds one winning path — typically the path that the heuristic function evaluates as most promising from the starting position. This path is winning, but it is not necessarily the most efficient path, the easiest path for a human to follow, or the path that best develops the strategic habits that transfer to future games. Solver paths frequently include moves that look actively counterproductive several steps before their benefit becomes apparent — moves that temporarily increase free cell occupancy, temporarily reduce foundation count, or temporarily destroy useful sequences — because the solver can see far enough ahead to value these regressive moves by their terminal position improvement. Human players who try to follow solver paths without understanding why each move is made often find the paths incomprehensible and lose confidence in their own judgment when their intuition disagrees with the solver's move. The correct use of solver analysis is not to follow the specific path but to understand which structural features of the position the solver's move is targeting — and to develop the position assessment skills that make those targets identifiable in real-time play.
Assuming that a solver that fails to find a solution confirms the deal is unwinnable. Not all solitaire solvers are complete — some use time limits or node count limits that terminate the search before exhausting all possible paths. A solver that terminates without finding a solution may have found a winning path if given more computation time; it has not confirmed that no winning path exists unless it exhaustively explored all paths and found none. Only a solver with a completeness guarantee — one that explores all reachable states and returns either a winning path or a confirmed exhaustion of all paths — can definitively confirm unwinnability. For practical purposes, the FreeCell unwinnable deals (and the unwinnable deal rates in other variants) in the statistics cluster were established by complete solvers with exhaustion guarantees. Player-accessible solvers online frequently lack this guarantee and their negative results should be treated as "no path found within the search budget" rather than "no path exists."
Treating solver win rates as achievable human targets. As covered in the FreeCell statistics guide, solver win rates approach the winnability floor — the proportion of deals that are intrinsically winnable by any legal sequence. Human players with expert strategy achieve 80–90% of FreeCell's ~99.975% winnability ceiling, 35–45% of Klondike's ~79–91% ceiling, and 20–30% of Forty Thieves' ~40–60% ceiling. The gap between solver ceiling and human expert performance is not primarily a knowledge gap — expert players know the correct principles — but a computational capacity gap: solvers can explore millions of positions per second while humans can evaluate perhaps five to ten per minute. Understanding this gap correctly means neither despairing that human play cannot match solver performance nor dismissing solver analysis as irrelevant to practical play. The correct relationship: solver analysis establishes what is achievable on the winnable deal population, and human strategy development aims to close as much of the human-solver gap as possible through better heuristics, better dead-end detection, and better backtracking — all of which are the skills described throughout this strategy cluster. For key card probability, see our key card probability guide; for the statistical foundation, see our FreeCell statistics guide.
Scorpion Solitaire illustrates the hidden information challenge that makes solver analysis of Klondike-family games expensive: three face-down columns at the start, combined with Spider-like same-suit build requirements, create a search space where the solver must branch over possible face-down card assignments while also managing suit consolidation requirements. A Scorpion player who develops the habit of mentally enumerating possible hidden card arrangements before each uncovering move is performing a simplified version of the probabilistic branching that Klondike-family solvers implement. Forty Thieves provides the sharpest demonstration of dead-end detection value: given its 40–60% unwinnable rate, an efficient dead-end detector that correctly identifies circular dependencies and stock exhaustion patterns early in the search dramatically reduces the computation wasted on unwinnable deals — which directly maps to the human habit of applying the three-pattern diagnostic before investing extended analysis time in stuck positions.
What is the best human strategy derived from how AI solvers work?Four solver-derived human habits produce the largest win rate improvements. The forced scan sequence (Foundation → Uncover → Pure build → Empty column → Stock) implements the solver's heuristic priority ordering. Undo-based hypothesis testing implements the solver's backtracking mechanism. Circular dependency checking implements the solver's dead-end detection pruning. And counter-intuitive path acceptance — being willing to make moves that look worse in the short term — implements the solver's willingness to explore low-immediate-heuristic moves that lead to high-terminal-heuristic positions. Together these four habits capture the essential architecture of a high-quality solitaire solver in a form that human players can execute in real time, without the millions-of-nodes-per-second computation that distinguishes solver from human performance.Which solitaire game is easiest for an AI solver to analyse?FreeCell is consistently easiest across all solver architectures because its complete information, near-100% winnability, and sufficient staging resources combine to produce a search problem where: the state is fully specified at every node (no hidden information branching); almost all deals have multiple winning paths (reducing the probability of exhausting the search without finding one); and the move-pruning rules are highly effective (free cell rotation constraints eliminate large fractions of the move graph quickly). Solvers routinely find FreeCell solutions within milliseconds to seconds. At the difficult end, Forty Thieves is hardest for unwinnable deal confirmation — its large state space and restricted build rules make the exhaustive path enumeration required to confirm unwinnability computationally expensive. Klondike's hidden information makes it hardest for exact winnability analysis, even though individual deal solution attempts are faster than Forty Thieves exhaustion.Can every solitaire game be solved by an AI solver with enough computing power?For complete-information variants (FreeCell, Yukon, Scorpion once all face-down cards are revealed), sufficient computing power solves any specific deal by exhaustive search — the search graph is finite and fully enumerable. For hidden-information variants (Klondike, Spider before all deals are triggered), the search graph is exponentially larger because of hidden card branching, and "enough computing power" is a harder threshold to define: unlimited power solves all cases, but the practical threshold for complete analysis of all possible Klondike deals is far beyond current hardware and represents an open problem in both theoretical computer science and practical solver engineering. The correct understanding is that AI solver capability is a spectrum — from near-instantaneous for easy FreeCell deals to computationally intractable for complete Klondike analysis — and human strategic improvement aims to close the gap between human and solver performance on the tractable end of that spectrum.
AI solvers can analyze various types of solitaire games, including Klondike, Spider, FreeCell, and Pyramid. Each game has its unique rules and strategies, which the solver must account for. Most solvers are designed to handle standard variations of these games, but some may also support custom rules or specific game formats. When choosing a solver, ensure it explicitly states compatibility with the solitaire variant you wish to analyze. Additionally, some solvers may offer insights into optimal strategies for each game type, helping players improve their skills.
To leverage an AI solver for improving your solitaire skills, start by inputting your game deals into the solver to analyze your moves. Pay attention to the suggested sequences and strategies provided by the solver. Take notes on the reasoning behind each move, especially in complex situations. After playing a game, compare your decisions with the solver's recommendations to identify mistakes or missed opportunities. Regularly practicing with the solver will help you understand the underlying strategies and enhance your decision-making skills in future games.
Yes, while AI solvers are powerful tools for analyzing solitaire games, they do have limitations. First, they may not always account for human factors such as intuition and risk assessment, which can lead to different strategies in real-time play. Additionally, some solvers may struggle with specific game variations or complex scenarios, potentially providing suboptimal advice. Furthermore, relying too heavily on solvers can hinder the development of your own strategic thinking. It's essential to balance solver use with personal practice to cultivate a deeper understanding of the game.