Why Reversi Attracted Computer Scientists

Reversi was one of the earliest board games targeted for computer play, and it wasn’t by accident. Several properties make it ideal for AI research:

  • Simple rules: The rules can be programmed in a few dozen lines of code
  • Small, fixed board: The 8×8 grid is large enough to be interesting but small enough for meaningful search
  • Fixed game length: A game lasts at most 60 moves (the number of empty squares at start), making the game tree bounded
  • Clear evaluation criteria: Mobility, stability, and corner control provide measurable positional features
  • Perfect information: Both players see the entire board at all times

These characteristics made reversi a natural testbed for game-playing algorithms — more complex than tic-tac-toe but more computationally approachable than chess.

Early Computer Players (1970s–1980s)

First Programs

The earliest reversi computer programs appeared in the late 1970s, coinciding with the Othello commercial boom and the first World Championship. These programs were simple by modern standards:

  • Limited search depth: Could look 3–5 moves ahead
  • Primitive evaluation: Judged positions mainly by disc count (ironically, the exact mistake that human beginners make)
  • No opening book: Started from scratch every game
  • No endgame solver: Played the endgame by the same approximate methods as the midgame

Despite these limitations, early programs were competent casual players and demonstrated the potential for AI in the game.

The Moor (1980)

The Moor, developed in the early 1980s, was one of the first reversi programs to achieve strong amateur-level play. It demonstrated that computer programs could outperform casual human players, though top competitors could still handle it comfortably.

Improving Evaluation Functions

The major breakthrough in early reversi AI was recognizing that disc count is a terrible evaluation metric. The best human players had always known this — disc maximizing is the #1 beginner mistake — but early programs had to learn it computationally.

By the mid-1980s, programs began using positional evaluation based on:

  • Mobility: How many legal moves does each player have?
  • Potential mobility: How many opponent frontier discs exist?
  • Corner ownership: Who controls the corners?
  • Stability: How many discs cannot be flipped?
  • Edge configurations: What’s the pattern along each edge?

These features aligned with what experienced human players had discovered through practice, but computers could assess them more precisely and consistently.

The Golden Age of Reversi AI (1990s)

Logistello

Logistello, developed by Michael Buro at the University of Alberta, was the program that definitively established computer supremacy in reversi. Logistello introduced several innovations:

Pattern-Based Evaluation

Instead of evaluating individual features (mobility, corners, etc.) independently, Logistello used pattern-based evaluation. It evaluated specific configurations of discs in specific board regions — for example, the exact pattern of all 8 squares along a single edge, or the 10-square pattern of a corner plus its adjacent squares.

This approach captured subtle positional details that simpler evaluations missed. Logistello had thousands of patterns, each with learned weight values, producing assessments far more accurate than any previous program.

Machine Learning

Logistello’s evaluation weights were trained using machine learning (specifically, a form of linear regression on game outcomes). The program played millions of games against itself, adjusting its evaluation to better predict which positions led to wins and losses.

This was an early application of what we now call machine learning in game AI — years before neural networks dominated the field.

Combined with alpha-beta pruning (the standard game tree search algorithm), Logistello could search 15–25 moves ahead in the midgame and solve the endgame perfectly from 24+ empty squares.

Opening Book

Logistello maintained an extensive opening book — a database of pre-analyzed positions covering the first 15–20 moves of every major opening. This meant the program played the opening phase almost instantly and with near-perfect accuracy.

Logistello vs. Murakami (1997)

The defining event in computer reversi was the 1997 match between Logistello and Takeshi Murakami, the reigning World Othello Champion.

Result: Logistello won 6–0.

This wasn’t close. Logistello dominated in openings, midgame, and endgame. Murakami, one of the strongest human players in history, was comprehensively outplayed.

The match proved that in reversi, the human era of competitive supremacy was over. Unlike chess, where the human-computer competition produced close matches for years (Kasparov vs. Deep Blue was 3.5–2.5), the reversi result was unambiguous.

Why was the reversi result so decisive?

  1. Endgame precision: Logistello could solve the endgame perfectly from 24 empty squares. Human players make errors in endgame counting — even the best occasionally miscalculate parity or disc swings
  2. Consistent evaluation: Computers don’t have bad days, fatigue, or nerves. Every position is evaluated with the same precision
  3. Opening book depth: The program’s opening preparation was more extensive than any human could memorize

Modern Engines (2000s–Present)

Post-Logistello Development

After Logistello proved computers could dominate human play, reversi engine development continued:

Edax

Edax, developed by Richard Delorme, became one of the most widely available strong reversi engines. It features:

  • Extremely efficient search algorithms
  • Perfect endgame solving from 24+ empty squares
  • A comprehensive opening book
  • Free availability for study and analysis

NTest

NTest, developed by Chris Welty, was another strong engine that contributed to opening theory through extensive self-play analysis.

Saio

Saio is a modern engine used by competitive players for game analysis and training. It represents the current state of the art in traditional (non-neural-network) reversi AI.

Neural Network Approaches

Following the success of neural networks in other games (notably AlphaGo and AlphaZero in Go and chess), researchers have applied similar techniques to reversi:

  • Deep learning models that evaluate positions based on raw board representations
  • Reinforcement learning approaches where programs learn entirely through self-play
  • Monte Carlo Tree Search (MCTS) combined with neural network evaluation

While these approaches produce strong play, traditional alpha-beta search with pattern-based evaluation remains competitive in reversi — the game’s relatively small search space means that brute-force search is more effective here than in games like Go.

How a Reversi Engine Works

A modern reversi engine combines four main components:

1. Opening Book

For the first 15–20 moves, the engine consults a pre-built database of analyzed positions. This book is created by deep analysis (sometimes to solved depth) of common openings. The engine plays these moves instantly and with near-perfect accuracy.

From roughly move 20 to move 40, the engine uses alpha-beta search:

  • Generate all legal moves
  • For each move, recursively evaluate the resulting position to a specified depth
  • Use pruning to skip positions that can’t improve the best-known option
  • Score leaf positions using the evaluation function

A strong engine searches 15–25 moves deep in the midgame, sometimes deeper with good pruning.

3. Evaluation Function

The evaluation function scores a position without searching further. Modern reversi evaluations consider:

  • Mobility: Legal move count for both players
  • Potential mobility: Frontier disc count (indicating future mobility)
  • Stability: Discs that can never be flipped
  • Corner ownership: Corner control and adjacent square patterns
  • Edge patterns: Complete edge configurations
  • Parity: Who has the favorable position for endgame parity

The function returns a numerical score — positive when the engine is ahead, negative when behind.

4. Endgame Solver

When approximately 20–24 empty squares remain, the engine switches from heuristic evaluation to perfect solving. It searches every possible continuation to the end of the game and determines the exact best move with certainty.

This is the area where computers most dramatically outperform humans. Human players estimate disc swings approximately; engines calculate exactly.

The Solved Game (2023)

6×6 Othello: Solved in 1993

The 6×6 variant of Othello was solved in 1993. The result: with perfect play, the second player (White) wins by 4 discs. This was computationally feasible because the 6×6 board has dramatically fewer positions than the standard 8×8.

8×8 Othello: Weakly Solved in 2023

In 2023, Japanese researcher Hiroki Takizawa announced that standard 8×8 Othello had been weakly solved. The result:

With perfect play from both sides, the game is a draw (33–31).

This means:

  • Neither player can force a win if both play perfectly
  • From the standard starting position with optimal play, the game ends with a narrow draw
  • Any deviation from perfect play by either player can be punished

The computation required enormous resources — it used extensive endgame databases and searched from the standard opening position through all critical variations.

What “weakly solved” means:

  • Strongly solved: The outcome is known from every possible position the game could reach
  • Weakly solved: The outcome is known from the standard starting position with both players playing optimally
  • Ultra-weakly solved: Only the theoretical outcome is known (win/loss/draw), but the actual perfect strategy isn’t fully mapped

Takizawa’s result is a weak solution — it proves the game is a draw from the standard opening but doesn’t necessarily provide perfect play from every reachable position.

Significance

Reversi joins checkers (solved in 2007 by Jonathan Schaeffer’s team at the University of Alberta) as one of the few classic board games whose theoretical outcome is known — see reversi vs checkers for a full comparison. The result confirms what strong players and computer programs had long suggested — reversi at perfect play is extremely balanced.

For human players, the practical impact is minimal. Humans are so far from perfect play that the theoretical draw result has no effect on actual games. But it represents a milestone in computational game theory.

AI’s Impact on Human Play

Computer analysis has profoundly influenced how humans study and play reversi:

Opening Theory

Opening databases created through engine analysis are now standard study material for competitive players. Engine analysis has refined opening evaluations, sometimes showing that positions previously thought to favor one side are actually balanced (or vice versa).

Endgame Training

Players use engines to practice endgame counting. A common training exercise: set up a position with 20 empty squares and try to find the perfect move, then check against the engine’s solution.

Game Review

After tournaments, competitive players analyze their games with engines to identify mistakes and understand what they missed. This is identical to how chess players use engines for post-game analysis.

Strategy Validation

Many strategic principles that human players discovered through experience — avoid X-squares, maximize mobility, fight for corners — have been validated by computer analysis. The engines confirm that these principles are correct, not just folk wisdom.