Landscapes and Other Art Forms.

Transcription

1 Landscapes and Other Art Forms. Darrell Whitley Computer Science, Colorado State University Permission to make digital or hard copies of part or all of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. Copyrights for third-party components of this work must be honored. For all other uses, contact the Owner/Author. Copyright is held by the owner/author(s).

2 Blind No More = GRAY BOX Optimization Darrell Whitley Computer Science, Colorado State University With Thanks to: Francisco Chicano, Gabriela Ochoa, Andrew Sutton and Renato Tinós

3 GRAY BOX OPTIMIZATION We can construct Gray Box optimization for pseudo-boolean optimization problems composed of M subfunctions, where each subfunction accepts at most k variables. Exploit the general properties of every Mk Landscape: m f(x) = f i (x) i=1 Which can be expressed as a Walsh Polynomial m W (f(x)) = W (f i (x)) i=1 Or can be expressed as a sum of k Elementary Landscapes k f(x) = ϕ (k) (W (f(x))) i=1

4 Walsh Example: MAXSAT

5 Walsh Example

6 Walsh Example

7 BLACK BOX OPTIMIZATION Don t wear a blind fold when walking about London if you can help it!

8 NK-Landscapes An Adjacent NK Landscape: n = 6 and k = 3. The subfunctions: f 0 (x 0, x 1, x 2 ) f 1 (x 1, x 2, x 3 ) f 2 (x 2, x 3, x 4 ) f 3 (x 3, x 4, x 5 ) f 4 (x 4, x 5, x 0 ) f 5 (x 5, x 0, x 1 ) These problems can be solved to optimality using Dynamic Programming.

9 Percent of Offspring that are Local Optima Using a Very Simple (Stupid) Hybrid GA: N k Model 2-point Xover Uniform Xover PX Adj 74.2 ± ± ± Adj 30.7 ± ± ± Adj 78.0 ± ± ± Adj 31.0 ± ± ± Rand 0.8 ± ± ± Rand 0.0 ± ± ± Rand 0.0 ± ± ± Rand 0.0 ± ± ±16.8

10 Number of partition components discovered N k Model Paired PX Mean Max Adjacent 3.34 ± Adjacent 5.24 ± Adjacent 7.66 ± Adjacent 7.52 ± Random 3.22 ± Random 2.41 ± Random 6.98 ± Random 2.46 ± Paired PX uses Tournament Selection. The first parent is selected by fitness. The second parent is selected by Hamming Distance.

11 Optimal Solutions for Adjacent NK 2-point Uniform Paired PX N k Found Found Found Percentage over 50 runs where the global optimum was Found in the experiments of the hybrid GA with the Adjacent NK Landscape.

12 Tunnelling Local Optima Networks NK Landscapes: Ochoa et al. GECCO 2015

13 NK and Mk Landscapes, P and NP

14 NK and Mk Landscapes, P and NP

15 But a Hybrid Genetic Algorithm is NOT how we should solve these NK Landscape Problems. We can exactly know the location of improving moves in constant time. No enumeration of neighbors is needed. Under well behaved conditions, we can exactly know the location of improving moves for r steps ahead in constant time.

16 Walsh Analysis Every n-bit MAXSAT or NK-landscape or P-spin problem is a sum of m subfunctions, f i : f(x) = m f i (x) The Walsh transform of f is a sum of the Walsh transforms of the individual subfunctions. W (f(x)) = i=1 m W (f i (x)) If m is O(n) then the number of Walsh coefficients is m 2 k = O(n). i=1

17 Mk Landscapes A General Model for all bounded Pseudo-Boolean Problems f(x) = m i = 1 fi (x, mask) f 1 f 2 f 3 f f 4 m

18 When 1 bit flips what happens? f(x) = m i = 1 fi (x, mask ) i f 1 f 2 f 3 f f 4 m flip

19 Constant Time Steepest Descent Assume we flip bit p to move from x to y p N(x). Construct a vector Score such that Score(x, y p ) = 2 1 bt x w b (x) b, p b In this way, all of the Walsh coefficients whose signs will be changed by flipping bit p are collected into a single number Score(x, y p ). The GSAT algorithm has done this for 23 years (Thanks to H. Hoos). NOTE: Hoos and Stützle have claimed a constant time result, but without proof. An average case complexity proof is required to obtain general constant time complexity results (Whitley 2013, AAAI). Also it does not matter if the problems are uniform random or not.

20 Best Improving and Next Improving moves Best Improving and Next Improving moves cost the same. GSAT uses a Buffer of best improving moves Buffer(best.improvement) =< M 10, M 1919, M 9999 > But the Buffer does not empty monotonically: this leads to thrashing. Instead uses multiple Buckets to hold improving moves Bucket(best.improvement) =< M 10, M 1919, M 9999 > Bucket(best.improvement 1) =< M 8371, M 4321, M 847 > Bucket(all.other.improving.moves) =< M 40, M 519, M 6799 > This improves the runtime of GSAT by a factor of 20X to 30X. The solution for NK Landscapes is only slightly more complicated.

21 The locations of the updates are obvious Score(y p, y 1 ) = Score(x, y 1 ) Score(y p, y 2 ) = Score(x, y 2 ) Score(y p, y 3 ) = Score(x, y 3 ) 2( w b(x)) b, (p 3) b Score(y p, y 4 ) = Score(x, y 4 ) Score(y p, y 5 ) = Score(x, y 5 ) Score(y p, y 6 ) = Score(x, y 6 ) Score(y p, y 7 ) = Score(x, y 7 ) Score(y p, y 8 ) = Score(x, y 8 ) 2( w b(x)) Score(y p, y 9 ) = Score(x, y 9 ) b, (p 8) b

22 The locations of the updates are obvious

23 What if we could look R Moves Lookahead? Consider R=3 Let Score(3, x, y i,j,k ) indicate we move from x to y i,j,k by flipping the 3 bits i, j, k. In general, we compute Score(r, x, y p ) when flipping r bits. f(y i ) = f(x) + Score(1, x, y i ) f(y i,j ) = f(y i ) + Score(1, y i, y j ) f(y i,j ) = f(x) + Score(2, x, y i,j ) f(y i,j,k ) = f(y i,j ) + Score(1, y i,j, y k ) f(y i,j,k ) = f(x) + Score(3, x, y i,j,k ) With thanks to Francisco Chicano!

24 Why Doesn t this exponentially EXPLODE??? f(y i,j,k ) = ((f(x) + Score(1, x, y i)) + Score(1, y i, y j)) + Score(1, y i,j, y k ) Score(3, x, y i,j,k ) = Score(2, x, y i,j) + Score(1, y i,j, y i,j,k ) If there is no Walsh Coefficient w i,j then Score(1, y i, y i,j ) = 0. Assume we have already moves of length shorter than 3. If there are no Walsh Coefficients linking i, j, k then Score(3, x, y i,j,k ) = 0.

25 The Variable Interaction Graph The Variable Interaction Graph Assume all distance 1 moves are taken. There cannot be a move flipping bits 4, 6, 9 that yields an improving move because there are no interactions and no Walsh coefficients.

26 Multiple Step Lookahead Local Search distance to optimum r=1 r=2 r=3 r=4 r=5 r= time in seconds In this figure, N = 12,000, k=3, and q=4. The radius is 1, 2, 3, 4, 5, 6. At r=6 the global optimum is found.

27 Steepest Descent on Moments Both f(x) and Avg(N(x)) can be computed with Walsh Spans. 3 f(x) = ϕ (z) (x) z=0 3 Avg(N(x)) = f(x) 1/d 2zϕ (p) (x) z=0 3 3 Avg(N(x)) = ϕ (z) (x) 1/N 2zϕ (z) (x) z=0 z=0 3 3 Avg(N(x)) = ϕ (z) (x) 2/N zϕ (z) (x) z=0 z=0

28 Steepest Descent on Moments Assume the vectors S p = Score p and Z p just map the change in the Walsh coefficients. Z p is computed using changes in zϕ (z) (x) ScoreAvg(y p, y 1 ) = ScoreAvg(x, y 1 ) ScoreAvg(y p, y 2 ) = ScoreAvg(x, y 2 ) ScoreAvg(y p, y 3 ) = ScoreAvg(x, y 3 ) + W eightedw alshupdate ScoreAvg(y p, y 4 ) = ScoreAvg(x, y 4 ) ScoreAvg(y p, y 5 ) = ScoreAvg(x, y 5 ) ScoreAvg(y p, y 6 ) = ScoreAvg(x, y 6 ) ScoreAvg(y p, y 7 ) = ScoreAvg(x, y 7 ) ScoreAvg(y p, y 8 ) = ScoreAvg(x, y 8 ) + W eightedw alshupdate ScoreAvg(y p, y 9 ) = ScoreAvg(x, y 9 )

29 Θ(1) Steepest Descent on Moments

30 What s (Obviously) Next? Local Search with r Move Lookahead PLUS Partition Crossover. Apply r Move Lookahead and Partition Crossover to MAX-kSAT. Use Deterministic Improving Moves. Use Deterministic Recombination.

31 What s (Obviously) Next? We can now solve 1 million variable NK-Landscapes to optimality in approximately linear time. (Paper submitted.)

32 What s (Obviously) Next? Put an End to the domination of Black Box Optimization. Wait for Tonight and Try to Take over the World.

33 THANK YOU Take Home Message: PROBLEM STRUCTURE MATTERS. Black Box Optimizers can never match the performance of an algorithm that efficiently exploits problem structure. But we need only a small amount of information: Gray Box Optimization. For Mk Landscapes, we can use Deterministic Moves and Deterministic Crossover.