Sporadic model building for efficiency enhancement of the hierarchical BOA

1. Sporadic model building for efficiency enhancement of the hierarchical BOA Genetic Programming and Evolvable Machines (2008) 9: 53-84 Martin Pelikan, Kumara Sastry, David E. Goldberg Summarized by Seok Ho-Sik

2. Core idea • Question • How could the time complexity of model building of multivariate probabilistic method be decreased ? • Observation • Evolution process for hBOA is composed of heavy computational process (structure learning) and light computational process (parameter learning). • Insight • Let’s reduce the computational burden by updating the structure of the probabilistic model once in every few iterations (generations).

3. Contributions • This paper describes and analyzes sporadic model building. • This paper also presents a dimensional model to provide a heuristic for scaling the structure-building period.

4. hBOA • Bayesian networks • Structure: an acyclicdirected graph with one node per variable and the edges corresponding to conditional probabilities (complexity: ). • Parameters: the conditional probabilities of each variable given the variables that this variable depends on (complexity: ). • To learn the structure of a Bayesian network with decision tree, a simple greedy algorithm is used. The greedy algorithm starts with an empty network, which is represented by single-node decision tree. Each iteration splits one leaf of any decision tree that improves the score of the network.

5. Sporadic model buildingSituations • The hierarchical Bayesian optimization (hBOA) algorithm can solve nearly decomposable and hierarchical optimization problems in a quadratic number of evaluations. • However, even quadratic performance may be insufficient for problems with many variables. • Q1: what is the computational bottleneck? A: structure learning of the Bayesian network  how could the computational burden be reduced? • Q2: how often should the structure be updated? A: once in every few iteration how to automatically set the period?

6. Sporadic model building (SMB) • Although the model structure does not remain the same throughout the run, the structure in consequent iterations of hBOA are often similar because consequent populations have similar structure with respect to the conditional dependencies and independencies. • SMB builds the structure once in every few iterations, while in remaining iterations only the parameters are updated. • Question: when to build the structure and when to only update the parameters? A: the structure-building period tsb the period when the structure is updated.

7. Effects of SMB on BOA • 1) Speed up of model building • Ideally, the speedup of building the model structure would be linearly proportional to tsb. • SMB may lead to an increase of the population size and the number of iteration. • 2) Slowdown of evaluation • SMB may lead to a decreased accuracy of the probabilistic models used to sample new candidates solutions  the population size N for building a sufficiently accurate model may increase with tsb. • Because of a less frequent adaptation of the model structure to the current population, SMB may slow down the convergence and the total number of iterations may increase with tsb.

8. Test problems • Dec -3: concatenated 3-bit deceptive function • The input string is partitioned into independent groups of 3 bits • A 3-bit deceptive function is applied to each group of 3 bits and the contributions of all deceptive functions are added to form the fitness. • Trap-5: concatenated 5-bit trap • Instead of 3-bit groups of Dec-3, 5-bit groups are considered. • hTrap: hierarchical Trap • Dec-3 and Trap-5 are separable problems of a fixed order. • hTrap constitute a problem that should be solved through hierarchical approach.

9. Experimental methodology • Problem size n = 30-210 with step 15 for dec-3 and trap-5. • n=9, 27,81, 243 for hTrap. • hBOA with SMB was tested for tsb from 1 to 20. • Each run was terminated either when the global optimum was found or when hBOA completed a large number of generations.

10. Results (1/4) • Speedup • Ideal case: the structure building speed-up would be the ratio of the number of times the structure was learned without SMB and the number of times the structure was learned with SMB. • SMB may lead to larger population size  real speed up = • Evaluation slowdown

11. Results (2/4)

12. Results (3/4)

13. Result (4/4) • SMB leads to a significant decrease of the asymptotic time complexity. • The optimal speedup obtained with SMB grows with problem size. • The facto by which SMB increases the number of evaluations is insignificant compared to the speedup of model building. • SMB with a constant structure-building period does not lead to an increase of the asymptotic complexity of hBOA with respect to the number of evaluations until convergence.

14. Dimensional model for separable problems of bounded order • The role of dimensional model: estimates the growth of tsb and the achieved speedup. • The basic ides • Assumption: the most of the subproblems have been captured by the current structure. • To compute the maximum number of iterations without structural updates to ensure that the model is rebuilt before the incorrectly modeled subproblems converge to non-optimal values. • To provide a worst-case analysis, the authors assume a situation where decision variables in the incorrectly modeled subproblems converge to non-optimal values.

15. Population sizing • The structure-building period should grow at most as fast as the square root of the number of subproblems. • Gambler’s ruin model: modeling the growth of inferior partial solutions • A gamble starts with some initial capital x0, an opponent with an initial capital N- x0. • The total number of tournaments to reach an absorbing state, tA (1) N: number of tournaments (2) (3) (4) (5) (6)

16. Ising spin glasses (1/3) • Question • Whether will SMB and the rule for scaling tsb achieve high speedups even on problems that are not trivially decomposable into subproblems of bounded order? • Ising spin glass • Each node i: spin si • Possible state of a spin: ±1 • Edge<i, j> and Ji,j: real value Ji,j defines the relationship between the two connected spins si and sj. • Energy: • Assuming the Boltzmann distribution of spin configurations.

17. Ising spin glasses (2/3) • The structure-building speedup grows with problem size. • For difficult problems, the evaluation slowdown with SMB becomes more significant than for separable problems, and that it grows with problem size.

18. Ising spin glasses (3/3)

19. Summary and conclusions • Sporadic model building leads to significant improvements of asymptotic time complexity of hBOA model building. • The slowdown due to an increase of the evalutions is insignificant compared to the model-building speedup. • Sporadic model building does not lead to an increase of the asymptotic complexity of hBOA with respect to the number of evaluations.

20. My questions on this paper • Statically determined tsb is not enough • Population sizing by gamble’s ruin model seems not to be sufficient for determining tsb. • Lower bound and upper bound should be provided. • Dynamically adjusted tsbwould be more superior than the introduced method. • The computational time for parameter learning is likely to be varied according to the population.