Regular PaperImproved sampling using loopy belief propagation for probabilistic model building genetic programming
Introduction
In the present paper, we introduce loopy belief propagation (LBP) in probabilistic model building GPs (PMBGPs) in order to generate the most probable solutions in sampling process. We call our novel method as POLE-BP.
Estimation of distribution algorithms (EDAs) are promising evolutionary algorithms and attract much attention from a lot of practical fields. EDAs optimize solution candidates represented by one dimensional arrays as well as Genetic Algorithms (GAs). Although EDA and GA employ the same chromosome representation, EDAs are different from GAs in the sense that EDAs generate new individuals by estimation of probabilistic models and sampling, whereas GAs generate them using genetic operators. EDAs can solve deceptive problems more efficiently than GAs by estimating dependencies between loci [27], which is one of the notable features of EDAs. Because of their effectiveness, many EDAs have been devised by incorporating many distinct statistical and machine learning approaches. Recently, EDAs using loopy belief propagation (LBP) as sampling were proposed in order to improve the sampling process [24], [21]. LBP approximately infers marginal and the highest joint probabilities with configurations, and has been applied to a wide range of real world problems [7], [6]. In EDAs, the individual with the highest joint probability in learned probabilistic models describes the models most and is often called as most probable solution (MPS). MPS is the individual which most reflects the learned models, and generation of it is important to take advantage of the models efficiently. However, traditional sampling methods used in EDAs, e.g. probabilistic logic sampling (PLS) [15] and Gibbs sampling, do not always generate MPS, and EDAs using only those samplings cannot make the best use of the models. In order to solve this problem, [24], [21] generate MPS by LBP in addition to traditional sampling and showed better search performance than existing methods using only traditional samplings (PLS or Gibbs sampling) in benchmark problems.
The estimation of distribution concept employed in EDAs has been applied to the optimization of tree structures, which is traditionally addressed using GP. GP optimizes tree structures using operators, such as crossover and mutation, as well as GA. Numerous improved genetic operators have been proposed because it is difficult to deal with tree structures using only these simple operators. EDAs for tree structures are often called as Genetic Programming-EDAs (GP-EDAs) [11] or Probabilistic Model Building GPs (PMBGPs) [32], and the present paper adopts the latter abbreviation throughout the paper. PMBGPs are broadly classified into two types. One type uses probabilistic context free grammar (PCFG) to represent distributions of promising solutions and learns production rule probabilities. The other type is a prototype tree based method, which converts trees to one dimensional arrays and applies EDAs to them. From the viewpoint of probabilistic models, the prototype tree-based method is essentially equivalent to EDAs and hence it can easily incorporate techniques devised in the field of EDA.
We propose POLE-BP [34], the novel prototype tree-based PMBGP with LBP. POLE-BP generates MPS at every generation in addition to normal samplings (i.e. PLS) and makes the optimal use of the learned probabilistic model. We compare our proposed method against existing methods on three benchmark problems: the problem with no dependencies between nodes (MAX problem), the deceptive problem (Deceptive MAX problem) and the problem with dependencies between nodes (Royal Tree Problem). From results of the experiments, we show that the proposed method competes with the existing method in the deceptive problem and beats the existing method in the problems with no deceptiveness from the point of the number of fitness evaluations to get an optimum solution. Moreover, we investigate behaviors of LBP in the context of PMBGP by observing fitness values and structures generated by LBP, and show reasons why the proposed method does not exhibit search performance improvement in deceptive problems whereas it does in other benchmark problems.
The present paper extends our prior work [34] by studying the effectiveness of LBP in detail. The remainder of the paper is organized as follows. Section 2 introduces related work. Section 3 explains details of the proposed method. Section 4 presents the experimental condition and results, which is followed by the discussion in Section 5. Finally Section 6 concludes the paper.
Section snippets
Related work
We introduce existing PMBGPs and methods using loopy belief propagation as sampling in this section.
The proposed method: POLE-BP
We briefly describe POLE-BP [34] in this section. POLE-BP is the first approach combining PMBGP and LBP. POLE-BP introduces LBP to the sampling process of Program Optimization with Linkage estimation (POLE) [12] and guarantees that population includes MPS, which is an individual having the highest joint probability, in each generation.
Prototype tree-based PMBGPs use more symbols than EDAs, which causes the following problems:
- 1.
require more population size and the number of evaluations;
- 2.
difficult
Experiments
In order to evaluate the search performance of the proposed method, we apply POLE-BP, POLE and simple GP (SGP) to three benchmark problems, MAX problem, deceptive MAX (DMAX) problem and royal tree problem and compare their performances. Common parameters in POLE-BP and POLE are described in Table 1, and parameters of SGP are described in Table 2. In POLE-BP, message passing schedule of loopy max-sum is that messages are sent from all factor nodes and variable nodes by turns. A termination
Discussion
In the previous section, for respective problems, we displayed the reduction of the average number of fitness evaluations by LBP and the behaviors of LBP in a PMBGP. Up to this point, we summarize the function of LBP in a PMBGP as the following: although LBP stimulates generation of building blocks, degree of the effectiveness differs dependent on the nature of problems. In the MAX problem, the best individual at each generation is apt to be generated by LBP. The MAX problem has no dependencies
Conclusion
We proposed POLE-BP, a variant of PMBGP that uses LBP for sampling to generate MPS at each generation. We compared the proposed POLE-BP against existing methods using the number of fitness function evaluations required as the evaluation criterion. Our experimental results showed that POLE-BPB׳s performance is statistically comparable with that of the existing methods for deceptive problems. Moreover, POLE-BP significantly outperformed other methods in non-deceptive problems. We analyzed the
References (38)
- et al.
An introduction and survey of estimation of distribution algorithms
Swarm Evol. Comput.
(2011) - S. Baluja, Population-based Incremental Learning: A Method for Integrating Genetic Search Based Function Optimization...
- P.A.N. Bosman, E.D. de Jong, Grammar Transformations in an EDA for genetic programming, in: GECCO 2004 Workshop...
- et al.
A Bayesian method for the induction of probabilistic networks from data
Mach. Learn.
(1992) - et al.
Toward understanding EDAs based on Bayesian networks through a quantitative analysis
IEEE Trans. Evol. Comput.
(2012) - R. Etxeberria, P. Larrañaga, Global optimization using Bayesian networks, in: Proceedings of 2nd Symposium on...
- A. Farinelli, A. Rogers, A. Petcu, N.R. Jennings, Decentralised coordination of low-power embedded devices using the...
- et al.
Efficient belief propagation for early vision
Int. J. Comput. Vis.
(2006) - C. Gathercole, P. Ross, S. Bridge, An adverse interaction between the crossover operator and a restriction on tree...
- G. Harik, Linkage Learning via Probabilistic Modeling in the ECGA, Technical Report,...
A Bayesian network approach to program generation
IEEE Trans. Evol. Comput.
Latent variable model for estimation of distribution algorithm based on a probabilistic context-free grammar
Trans. Evol. Comput.
A review on probabilistic graphical models in evolutionary computation
J. Heuristics
Cited by (4)
Computing budget allocation in multi-objective evolutionary algorithms for stochastic problems
2018, Swarm and Evolutionary ComputationCitation Excerpt :Researchers have developed different noise handling techniques to solve stochastic problems, aiming to improve the accuracy and efficiency of the algorithms. For example, a probabilistic method to improve sampling using loopy belief propagation for probabilistic model building genetic programming is described in [5]. Population statistics based re-sampling technique is introduced in [6] with the particle swarm optimization algorithm to solve stochastic optimization problems.
Evolutionary approach to machine learning and deep neural networks: Neuro-evolution and gene regulatory networks
2018, Evolutionary Approach to Machine Learning and Deep Neural Networks: Neuro-Evolution and Gene Regulatory NetworksGE and semantics
2018, Handbook of Grammatical Evolution