Genetic transfer learning

doi:10.1016/j.eswa.2010.03.019

Expert Systems with Applications

Volume 37, Issue 10, October 2010, Pages 6997-7002

https://doi.org/10.1016/j.eswa.2010.03.019 Get rights and content

Abstract

Transfer learning is a method which aims to improve “related” tasks performance. Transfer learning tries to use information gained from related tasks solutions to improve performance of learning strategy. Transfer learning addresses the problem of how to utilize plenty of labeled data in a source domain to solve related but different problems in a target domain, even when the training and testing problems have different distributions or features (Pan, Kwok, & Yang, 2008). In this paper we have used transfer learning to improve performance of genetic algorithms.

Introduction

Traditional machine learning techniques try to solve one problem at a time. And some of the complex problems take a lot of time to solve by classic machine learning techniques. Also if there is not enough information about the problem it’s sometimes difficult to get the exact solution. Fig. 1 shows the difference of learning process between traditional machine learning techniques and transfer learning techniques. Traditional machine learning techniques try to learn each task from scratch, while transfer learning techniques try to transfer knowledge from previous tasks to a new target task with a few training data (Pan & Yang, 2008). In real life a human can solve complex problems easily by using his/her past experiences in the problem domains. Transfer learning is a humanlike learning strategy which aims to help problem solving and learning mechanisms by using information gained from previously solved related tasks. Transfer learning is used in many learning methods like neural networks (Murre, 1995, Pratt, 1993), markov logic networks (Mihalkova and Huynh, 2007, Mihalkova and Mooney, 2006), text categorization (Gupta & Ratinov, 2008), web page classification (Ling et al., 2008).

The key problem of transfer learning is task relatedness. Determining whether the tasks are related or not which the information will be transferred among is very important or how they are correlated. And decide transfer or not information between two tasks, how many information to transfer and how the information will be transferred.

Transfer learning is first used to set initial weight of a “target” neural network by using another, trained, same architecture (e.g. same size input and same size output) “source” neural network (Pratt, 1993). When the information gained from source network used directly (e.g. weights) it performs worse in target network than random initial weights because, weights in source network was too big and reduced very slowly by back-propagation. So gained information from source network is reduced by the benefit, which tested by target networks training data.

Another transfer learning example is markov logic networks, which consist of a set of weighted formulae and provides a way of softening first-order logic by making situations, in which not all formulae are satisfied, less likely but not impossible (Mihalkova & Huynh, 2007). In this working three real-world relational domains IMDB, UW-CSE, and WebKB are used. The aim of the working is training faster by transferring obtained rules form cross domains.

Transfer learning is also used in classification problems. Tasks sampled from the environment are used to improve classification performance on future tasks. We consider situations in which the tasks can be divided into groups. Tasks within each group are related by sharing a low dimensional representation, which differs across the groups (Argyriou, Maurer, & Pontil, 2008).

Transfer learning mostly used in reinforcement learning and lifelong learning. One of the fundamental working is “Lifelong Robot Learning” (Thrun & Mitchell, 1995).

More detailed information about transfer learning can be found in (Pan & Yang, 2008).

All workings above acquired good performance improvements. So we would like to try if we can acquire good performance improvements in other learning strategies like genetic algorithms.

Section snippets

Genetic transfer

Genetic programming is a domain-independent problem solving approach in which computer programs are evolved to solve, or approximately solve, problems. Genetic programming is based on the Darwinian principle of reproduction and survival of the fittest and analogs of naturally occurring genetic operations such as crossover (sexual recombination) and mutation (Koza, 1998). More detailed information can be found in (Gen and Cheng, 2000, Goldberg, 1989).

The aim in genetic transfer is, like the

Experimental results

In our experiments we tried two different “individual pool” creation methods. All distance and difference values are normalized values by the Eqs. (7), (8). Because values of these variables are depend on individual length $Normalized difference = \frac{difference}{max_value}$ $Normalized distance = \frac{distance}{max_value}$

–
Single transfer (S.T.): Single transfer pool creation method is used to create “individual pool” by only one running of source function, so size of individual pool is 100 × 3 = 300 for iteration count

Conclusion and future work

The “goodness” value is the probability that genetic transfer algorithm will have better performance than classic genetic algorithm, calculated by 10 independent runs. And similarly “difference” value is the answer for question “how much genetic transfer algorithm is better than classic genetic algorithm?” calculated by best fitness values of 10 independent runs. As it seen in experimental result, genetic transfer algorithm is a good choice when you have two similar optimization problems. In

References (15)

V.J. Milenkovic et al.
Translational polygon containment and minimal enclosure using mathematical programming
International Transactions in Operational Research
(1999)
Argyriou, A., Maurer, A., & Pontil, M. (2008). An algorithm for transfer learning in a heterogeneous environment. In...
M. Gen et al.
Genetic algorithms and engineering optimization
(2000)
D.E. Goldberg
Genetic algorithms in search, optimization and machine learning
(1989)
Gupta, R., & Ratinov, L. (2008). Text categorization with knowledge transfer from heterogeneous data sources. In...
Hassan Mahmud, M. M. (2007). On universal transfer learning. In Proceedings of the 18th international conference on...
J. Koza
Genetic programming
Encyclopedia of Computer Science and Technology
(1998)

There are more references available in the full text version of this article.

Cited by (55)

Transfer learning and its extensive appositeness in human activity recognition: A survey
2024, Expert Systems with Applications
In this competitive world, the supervision and monitoring of human resources are primary and necessary tasks to drive context-aware applications. Advancement in sensor and computational technology has cleared the path for automatic human activity recognition (HAR). First, machine learning and later deep learning play a cardinal role in this automation process. Classical machine learning approaches follow the hypothesis that the training, validation, and testing data belong to the same domain, where data distribution characteristics and the input feature space are alike. However, during real-time HAR, the above hypothesis does not always true. Transfer learning helps in an extended manner to transfer the required knowledge among heterogeneous data of various activities. To display the hierarchical advancements in transfer learning-enhanced HAR, we have shortlisted the 150 most influential works and articles from 2014–2021 based on their contribution, citation score, and year of publication. These selected articles are collected from IEEE Xplore, Web of Science, and Google Scholar digital libraries. We have also analyzed the statistical research interest related to this topic to substantiate the significance of our survey. We have found a significant growth of 10% in research publications related to this domain every year. Our survey provides a unique classification model to delineate the diversity in transfer learning-based HAR. This survey delves into the world of HAR datasets, exploring their types, specifications, advantages, and limitations. We also examine the steps involved in HAR, including the various transfer learning techniques and performance metrics, as well as the computational complexity associated with these methods. Additionally, we identify the challenges and gaps in HAR related to transfer learning and provide insights into future directions for researchers in this field. Based on the survey findings, researchers prefer the inductive transfer method, feature learning transfer mode, and cross-action transfer domain more over others due to their superior performance, with respective popularity scores of 55%, 40.8%, and 50.2%. This review aims to equip readers with a comprehensive understanding of HAR and transfer learning mechanisms, while also highlighting areas that require further research.
Migration-based algorithm library enrichment for constrained multi-objective optimization and applications in algorithm selection
2023, Information Sciences
It is of necessity to select appropriate optimization algorithms from an algorithm library due to the universality of constrained multi-objective optimization problems and the suitability of intelligent optimization algorithms, which requires a rich optimization algorithm library. This paper proposes a migration-based method of enriching the algorithm library for constrained multi-objective optimization problems. After calculating the similarity between problems based on their landscape features, the proposed method calculates the migration probabilities of intelligent optimization algorithms solving similar problems based on the performance of each algorithm and the similarity between problems. According to the redundancy and compatibility of components, the algorithms with large migration probabilities enrich the algorithm library for solving the current problem. Based on the enhanced algorithm library, a Softmax regression model is trained to generate an optimal intelligent algorithm to solve the current problem. The proposed method is applied to solve a series of constrained multi-objective optimization benchmark problems and the operation optimization problems of an integrated coal mine energy system, and the experimental results verify its effectiveness and feasibility.
Transfer learning in optimization: Interpretable self-organizing maps driven similarity indices to identify candidate source functions
2023, Expert Systems with Applications
In the design evolution of a product, designers often require solving similar functions repeatedly across different designs. These functions are usually related to each other and typically share topology, common features and physics. Thus, solving one function referred to as a source that characterizes a problem can yield knowledge that can be reused to solve other related function referred to as target. Re-purposing such shared knowledge, especially in cases of complex optimization, aids in faster convergence, more accurate solutions, and reduced computational costs, among others. The concept of transfer learning (TL) is built on this notion of passing the gained knowledge between related problems to lessen the algorithmic and/or modeling complexities. Transfer of knowledge between source and target that are not related leads to negative transfer circumstances, where the algorithm’s performance degrades on transferring. Hence, identifying a similar source to share the knowledge is of fundamental importance in transfer learning approaches. Literature has often skipped this step of identifying related or similar functions by artificially constructing functions or assuming apriori that the considered functions are similar to each other. Current work proposes to use interpretable self-organizing map and an image comparison technique to quantify the topological similarity between the source and the target. Metrics such as mean squares, structural similarity index, and Cosine similarity are used to quantify the level of similarity mathematically. The proposed approach is implemented on a suite of benchmark analytical functions with varying order, complexity, and dimensions, engineering examples, and real-world application functions. It is demonstrated that the proposed approach is able to identify appropriate source function for a given target, even when they are of varying dimensions. Results of engineering examples show that functions representing problems with similar physics are identified correctly. Hence, the proposed approach can be used to identify appropriate source functions for a given target, permitting transfer learning and thus accelerating convergence, and reducing computational cost.
Cross-domain feature selection and diagnosis of oil and gas pipeline defects based on transfer learning
2023, Engineering Failure Analysis
Citation Excerpt :
To address cross-domain diagnosis problems, TL methods have been proposed by many researchers [21,22]. Combining TL with the feature selection method or ML method can relax the restriction that test and training data must come from the same domain [23]. In many TL-based feature selection methods, an attempt is made to create a new feature projection space, and projected target data can be classified by traditional ML algorithms [24,25].
Developing intelligent pipeline defect diagnosis technology is of great significance to ensure the efficiency and safety of pipeline transportation. However, a challenge in real-world tasks is the inadaptability and instability of these technologies for processing data from different distributions. Therefore, this paper proposes an intelligent transfer defect diagnosis framework that combines a hybrid feature selection method and transfer learning (TL) technology. The hybrid feature selection method consists of two cross-domain filtering feature selection methods and feature subset evaluation. A Monte Carlo simulation wrapper is used to obtain the best transfer feature subset. The obtained optimal feature subset is fed to the feature adaptation-based classification method for defect diagnosis. We used two different datasets as the target domain and source domain and simulated two transfer scenarios to verify the framework. Experimental results demonstrate that the proposed diagnosis framework achieves high accuracy and effectiveness in pipeline defect diagnosis cross-domain transfer tasks. This will facilitate pipeline maintenance engineers in solving the problem of pipeline defect diagnosis and reasonably arranging pipeline maintenance.
A deep hybrid transfer learning-based evolutionary algorithm and its application in the optimization of high-order problems
2023, Soft Computing
Knowledge Transfer Genetic Programming With Auxiliary Population for Solving Uncertain Capacitated Arc Routing Problem
2023, IEEE Transactions on Evolutionary Computation

View all citing articles on Scopus

View full text

Genetic transfer learning

Abstract

Introduction

Section snippets

Genetic transfer

Experimental results

Conclusion and future work

International Transactions in Operational Research

Genetic algorithms and engineering optimization

Genetic algorithms in search, optimization and machine learning

Genetic programming

Encyclopedia of Computer Science and Technology