270 likes | 470 Views
IEEE CEC 2006. Product Geometric Crossover for the Sudoku Puzzle. Alberto Moraglio, Julian Togelius & Simon Lucas. Contents. Geometric Crossover and Product Geometric Crossover Design of Geometric Crossover for the Sudoku Puzzle Experimental Results and Conclusions. I. Geometric Crossover.
E N D
IEEE CEC 2006 Product Geometric Crossover for the Sudoku Puzzle Alberto Moraglio, Julian Togelius & Simon Lucas
Contents • Geometric Crossover and Product Geometric Crossover • Design of Geometric Crossover for the Sudoku Puzzle • Experimental Results and Conclusions
y x Geometric Crossover • Line segment • A binary operator GX is a geometric crossover if all offspring are in a segment between its parents. • Geometric crossover is dependent on the metric
B 1 0 1 1 0 1 X A 2 B 1 1 0 1 1 3 A X 0 1 0 1 1 Geometric Crossover • The traditional n-point crossover is geometric under the Hamming distance. H(A,X) + H(X,B) = H(A,B)
Many Recombinations are Geometric • Traditional Crossover extended to multary strings • Recombinations for real vectors • PMX, Cycle Crossovers for permutations • Homologous Crossover for GP trees • Ask me for more examples over a coffee!
Being geometric crossover is important because…. • We know how the search space is searched by geometric crossover for any representation: convex search • We know a rule-of-thumb on what type of landscapes geometric crossover will perform well: “smooth” landscape • This is just a beginning of general theory, in the future we will know more!
Product Geometric Crossover • GX1:AxAA geometric under d1 • GX2:BxB B geometric under d2 • A product crossover of GX1 and GX2 is an operator defined on the cartesian product of their domains PGX:(A,B)x(A,B)(A,B) that applies GX1on the first projection and GX2 on the second projection. GX1 and GX2 do not need to be independent. • Theorem: PGX is a geometric crossover under the distance d = d1+d2
Properties of Product Geometric Crossover • It is a simple and general method to build more complex geometric crossovers from simple geometric crossovers • Multi-crossover: same representation, same crossover n times • Hybrid crossover: same representation, different crossover for each projection • Hybrid representation: different representation (and crossover) for each projection • No independence required: base crossovers do not need to be independent
The Sudoku Game Fill in the grid so that every row,every column, and every 3x3 boxcontains the digits 1 through 9
Constraints It is a constraint-satisfaction problem with 4 types of constraints: • Fixed Elements • Rows are permutations • Columns are permutations • Boxes are permutations
Computational Complexity • The general Sudoku puzzle is based on a (n^2)x(n^2) grid • The problem is NP-Complete • Relaxation (3 constraints) • Latin square completion: (1+2+3) NP-Hard • Sudoku puzzle generator: (2+3+4) Polynomial? • Initialisation problem: (1+2+4 or 1+3+4) NP-Hard? • Relaxation (2 constraints): Polynomial!
Geometric Design • Look at the problem and build a nice fitness landscape (= fitness function + distance) • the smaller search space the better • the smoother landscape the better • Pick genetic operators that match the landscape: mutation and crossover should be geometric under the distance chosen
Soft & Hard Constraints • Hard constraints: all feasible solutions must respect them. Search operators take feasible solutions and produce feasible solutions • Soft constraints: level of fulfillment is the fitness of a solution • More than one combination of soft and hard constraints available!
Restricted Hamming space • Hard constraint: fixed positions • Soft constraints: permutations on rows, columns and boxes • Distance: Hamming distance between grids • Feasible Mutation: change any non-fixed element • Feasible Crossover: traditional crossover over the vector obtained by joining the rows of the grid
Row-swap space • Hard constraints: fixed positions and permutations on rows • Soft constraints: permutations on columns and boxes • Distance: sum of swap distances between paired rows (row-swap distance) • Feasible mutation: swap two non-fixed elements in a row
Geometric crossovers for row-swap space • Row-wise PMX and row-wise cycle crossover • Feasibility: • Row permutation: simple PMX and cycle crossovers recombine permutations and produce permutations • Fixed elements: they both preserve fixed positions in the parents • Geometricity: • Known: simple PMX and cycle crossovers are geometric under swap distance • For the product geometric theorem: row-wise PMX and row-wise cycle crossovers are geometric under row-swap distance
Fitness Function • Fitness: level of fulfilment of soft constraints • Fitness to maximize: • Sum of unique elements in each row, plus, • Sum of unique elements in each column, plus, • Sum of unique elements in each box • For a 9x9 grid the fitness corresponding to a fully correct grid is 243
Smooth Fitness Landscapes • Restricted Hamming Space: • a single element change affects the current fitness of -1, 0 or +1 for its row, for its column and for its box. Absolute maximum total change in fitness for a single change is 3 • Row-swap space: • A single swap in a row affects the current fitness of 0 for its row, between -2 and +2 for the columns touched, and the same for the boxes touched. The absolute maximum total change in fitness for a single swap in a row is 4 • Maximum delta fitness: • Max fitness for 9x9 grid: 243 • Min fitness for 9x9 grid: 27 • Max delta fitness in the landscape: 243 – 27=216 • Index of smoothness: • Change in fitness at distance one divided maximum change in fitness • 0: perfectly smooth landscape, 1: max and min fitness are neighbours • Index for Restricted Hamming Space: 3/216 • Index for Row-swap Space: 4/216 • Both Fitness Landscapes are very smooth!
Prediction! • Both fitness landscapes are very smooth so geometric crossovers and mutations associated with both spaces should work well • Advantages of the row-swap search space: • it is much smaller because it restricts the search to feasible rows • The restriction includes the optimum grid and prunes only grids with lower fitness • Bet: Row-swap operators will win!
Results summary • Crossovers based on row-swap space better than those based on hamming space • Crossover (with mutation) better than hill climbers • Many more experiments in the paper! • Future work: smartsquare + crossover
Conclusions • Extended the geometric crossover with the notion of Product Geometric Crossover • Product geometric crossover for Sudoku • Designed geometric crossovers to deal naturally with constraints • New geometric crossovers for the entire grid by using simple geometric crossover for each rows • The associated distance has allowed us to analyse the crossover fitness landscape and predict that the crossovers will perform well • Extensive experimental results confirm that the crossovers designed perform well