God's algorithm

The notion applies to puzzles that can assume a finite number of "configurations", with a relatively small, well-defined arsenal of "moves" that may be applicable to configurations and then lead to a new configuration. Solving the puzzle means to reach a designated "final configuration", a singular configuration, or one of a collection of configurations. To solve the puzzle a sequence of moves is applied, starting from some arbitrary initial configuration.

An algorithm can be considered to solve such a puzzle if it takes as input an arbitrary initial configuration and produces as output a sequence of moves leading to a final configuration (if the puzzle is solvable from that initial configuration, otherwise it signals the impossibility of a solution). A solution is optimal if the sequence of moves is as short as possible. The highest value of this, among all initial configurations, is known as God's number,^[3] or, more formally, the minimax value.^[4] God's algorithm, then, for a given puzzle, is an algorithm that solves the puzzle and produces only optimal solutions.

Some writers, such as David Joyner, consider that for an algorithm to be properly referred to as "God's algorithm", it should also be practical, meaning that the algorithm does not require extraordinary amounts of memory or time. For example, using a giant lookup table indexed by initial configurations would allow solutions to be found very quickly, but would require an extraordinary amount of memory.^[5]

Instead of asking for a full solution, one can equivalently ask for a single move from an initial but not final configuration, where the move is the first of some optimal solution. An algorithm for the single-move version of the problem can be turned into an algorithm for the original problem by invoking it repeatedly while applying each move reported to the present configuration, until a final one is reached; conversely, any algorithm for the original problem can be turned into an algorithm for the single-move version by truncating its output to its first move.

Well-known puzzles fitting this description are mechanical puzzles such as Rubik's Cube, the Tower of Hanoi, and the 15 puzzle. The one-person game of peg solitaire is also covered, as well as many logic puzzles, such as the missionaries and cannibals problem. These have in common that they can be modeled mathematically as a directed graph, in which the configurations are the vertices, and the moves are the arcs.

The Fifteen puzzle can be solved in 80 single-tile moves^[6] or 43 multi-tile moves^[7] in the worst case. For its generalization the n-puzzle, the problem of finding an optimal solution is NP-hard,^[8] so it is not known whether there is a practical God's algorithm.

For the Towers of Hanoi puzzle, a God's algorithm is known for any given number of disks. The number of moves increases exponentially with the number of disks (${\displaystyle 2^{n}-1}$ ).^[9]

 

An algorithm to determine the minimum number of moves to solve Rubik's Cube was published in 1997 by Richard Korf.^[10] While it had been known since 1995 that 20 was a lower bound on the number of moves for the solution in the worst case, Tom Rokicki proved in 2010 that no configuration requires more than 20 moves.^[11] Thus, 20 is a sharp upper bound on the length of optimal solutions. Mathematician David Singmaster had "rashly conjectured" this number to be 20 in 1980.^[4]

Some well known games with a very limited set of simple well-defined rules and moves have nevertheless never had their God's algorithm for a winning strategy determined. Examples are the board games chess and Go.^[12] Both these games have a rapidly increasing number of positions with each move. The total number of all possible positions, approximately 5×10^44[13] for chess and 10¹⁸⁰ (on a 19×19 board) for Go,^[14] is much too large to allow a brute force solution with current computing technology (compare the now solved, with great difficulty, Rubik's Cube at only about 4.3×10¹⁹ positions^[15]). Consequently, a brute force determination of God's algorithm for these games is not possible. While chess computers have been built that are capable of beating even the best human players, they do not calculate the game all the way to the end. Deep Blue, for instance, searched only 11 moves ahead (counting a move by each player as two moves), reducing the search space to only 10¹⁷.^[16] After this, it assessed each position for advantage according to rules derived from human play and experience.

Even this strategy is not possible with Go. Besides having hugely more positions to evaluate, no one so far has successfully constructed a set of simple rules for evaluating the strength of a Go position as has been done for chess, though neural networks trained through reinforcement learning can provide evaluations of a position that exceed human ability.^[17] Evaluation algorithms are prone to make elementary mistakes^[18] so even for a limited look ahead with the goal limited to finding the strongest interim position, a God's algorithm has not been possible for Go.

On the other hand, draughts (checkers) has long been suspected of being "played out" by its expert practitioners.^[19] In 2007 Schaeffer et al. proved this to be so by calculating a database of all positions with ten or fewer pieces, providing a God's algorithm for all end games of draughts which was used to prove that all perfectly played games of draughts end in a draw.^[20] However, draughts with only 5×10²⁰ positions^[21] and even fewer, 3.9×10¹³, in the database,^[22] is a much easier problem to solve –of the same order as Rubik's cube.

The magnitude of the set of positions of a puzzle does not entirely determine whether a God's algorithm is possible. The already solved Tower of Hanoi puzzle can have an arbitrary number of pieces, and the number of positions increases exponentially as ${\displaystyle 3^{n}}$ . Nevertheless, the solution algorithm is applicable to any size problem, with a running time scaling as ${\displaystyle 2^{n}}$ .^[23]

• Oracle machine
• Divine move (game of Go)
• Proofs from THE BOOK
• Rubik's Cube group
• Solved game

• Baum, Eric B., What is Thought?, MIT Press, 2004 ISBN 0262025485.
• Davis, Darryl N.; Chalabi, T.; Berbank-Green, B., "Artificial-life, agents and Go", in Mohammadian, Masoud, New Frontiers in Computational Intelligence and its Applications, pp. 125–139, IOS Press, 2000 ISBN 9051994761.
• Fraser, Rober (ed); Hannah, W. (ed), The Draught Players' Weekly Magazine, vol. 2, Glasgow: J H Berry, 1885.
• Joyner, David (2002). Adventures in Group Theory. Johns Hopkins University Press. ISBN 0-8018-6947-1.
• Moore, Cristopher; Mertens, Stephan, The Nature of Computation, Oxford University Press, 2011 ISBN 0191620807.
• Rothenberg, Gadi, Catalysis, God's Algorithm, and the Green Demon, Amsterdam University Press, 2009 ISBN 9056295896.
• Schaeffer, Jonathan; Burch, Neil; Björnsson, Yngvi; Kishimoto, Akihiro; Müller, Martin; Lake, Robert; Lu, Paul; Sutphen, Steve (14 September 2007). "Checkers Is Solved" (PDF). Science. 317 (5844): 1518–1522. Bibcode:2007Sci...317.1518S. doi:10.1126/science.1144079. ISSN 0036-8075. PMID 17641166.
• Singmaster, David, Notes on Rubik's Magic Cube, Penguin, 1981 ISBN 0-907395-00-7.
• Singmaster, David, "The educational value of the Hungarian 'Magic Cube'", Proceedings of the Fourth International Congress on Mathematical Education, held in Berkeley, California, 10–16 August 1980, pp. 307–312, Birkhauser Boston Inc, 1983 ISBN 978-0-8176-3082-9.