Combinations of Estimation of Distribution Algorithms and Other Techniques

This paper summaries our recent work on combining estimation of distribution algorithms （EDA） and other techniques for solving hard search and optimization problems： a） guided mutation, an offspring generator in which the ideas from EDAs and genetic algorithms are combined together, we have shown that an evolutionary algorithm with guided mutation outperforms the best GA for the maximum clique problem, b） evolutionary algorithms refining a heuristic, we advocate a strategy for solving a hard optimization problem with complicated data structure, and c） combination of two different local search techniques and EDA for numerical global optimization problems, its basic idea is that not all the new generated points are needed to be improved by an expensive local search.

Learning to select the recombination operator for derivative-free optimization

Extensive studies on selecting recombination operators adaptively,namely,adaptive operator selection(AOS),during the search process of an evolutionary algorithm(EA),have shown that AOS is promising for improving EA's performance.A variety of heuristic mechanisms for AOS have been proposed in recent decades,which usually contain two main components:the feature extraction and the policy setting.The feature extraction refers to as extracting relevant features from the information collected during the search process.The policy setting means to set a strategy(or policy)on how to select an operator from a pool of operators based on the extracted feature.Both components are designed by hand in existing studies,which may not be efficient for adapting optimization problems.In this paper,a generalized framework is proposed for learning the components of AOS for one of the main streams of EAs,namely,differential evolution(DE).In the framework,the feature extraction is parameterized as a deep neural network(DNN),while a Dirichlet distribution is considered to be the policy.A reinforcement learning method,named policy gradient,is used to train the DNN.As case studies,the proposed framework is applied to two DEs including the classic DE and a recently-proposed DE,which result in two new algorithms named PG-DE and PG-MPEDE,respectively.Experiments on the Congress of Evolutionary Computation(CEC)2018 test suite show that the proposed new algorithms perform significantly better than their counterparts.Finally,we prove theoretically that the considered classic methods are the special cases of the proposed framework.

Learning to sample initial solution for solving 0-1 discrete optimization problem by local search

Local search methods are convenient alternatives for solving discrete optimization problems(DOPs).These easy-to-implement methods are able to find approximate optimal solutions within a tolerable time limit.It is known that the quality of the initial solution greatly affects the quality of the approximated solution found by a local search method.In this paper,we propose to take the initial solution as a random variable and learn its preferable probability distribution.The aim is to sample a good initial solution from the learned distribution so that the local search can find a high-quality solution.We develop two different deep network models to deal with DOPs established on set(the knapsack problem)and graph(the maximum clique problem),respectively.The deep neural network learns the representation of an optimization problem instance and transforms the representation to its probability vector.Experimental results show that given the initial solution sampled from the learned probability distribution,a local search method can acquire much better approximate solutions than the randomly-sampled initial solution on the synthesized knapsack instances and the Erd?s-Rényi random graph instances.Furthermore,with sampled initial solutions,a classical genetic algorithm can achieve better solutions than a random initialized population in solving the maximum clique problems on DIMACS instances.Particularly,we emphasize that the developed models can generalize in dimensions and across graphs with various densities,which is an important advantage on generalizing deep-learning-based optimization algorithms.

Demonstration of a diode-pumped dual-wavelength metastable krypton laser

Diode-pumped rare gas lasers are potential candidates for high-energy and high-beam quality laser systems.Currently,most investigations are focused on metastable Ar lasers.The Kr system has the unique advantages of higher quantum efficiency and lower discharge requirements for comparison.In this paper,a diode-pumped metastable Kr laser was demonstrated for the first time.Using a repetitively pulsed discharge at a Kr/He pressure of up to approximately1500 Torr,metastable Kr atoms of more than 10^(13)cm^(-3)were generated.Under diode pumping,the laser realized a dual-wavelength output with an average output power of approximately 100 mW and an optical conversion efficiency of approximately 10% with respect to the absorbed pump power.A kinetics study involving population distribution and evolution was conducted to analyze the laser performance.

MSIsensor-pro: Fast, Accurate, and Matched-normal-sample-free Detection of Microsatellite Instability

Microsatellite instability(MSI)is a key biomarker for cancer therapy and prognosis.Traditional experimental assays are laborious and time-consuming,and next-generation sequencingbased computational methods do not work on leukemia samples,paraffin-embedded samples,or patient-derived xenografts/organoids,due to the requirement of matched normal samples.Herein,we developed MSIsensor-pro,an open-source single sample MSI scoring method for research and clinical applications.MSIsensor-pro introduces a multinomial distribution model to quantify polymerase slippages for each tumor sample and a discriminative site selection method to enable MSI detection without matched normal samples.We demonstrate that MSIsensor-pro is an ultrafast,accurate,and robust MSI calling method.Using samples with various sequencing depths and tumor purities,MSIsensor-pro significantly outperformed the current leading methods in both accuracy and computational cost.MSIsensor-pro is available at https://github.com/xjtu-omics/msisensor-pro and free for non-commercial use,while a commercial license is provided upon request.

Aberrantly high activation of a FoxM1-STMN1 axis contributes to progression and tumorigenesis in FoxMl-driven cancers

Fork-head box protein M1(FoxM1)is a transcriptional factor which plays critical roles in cancer development and progression.However,the general regulatory mechanism of FoxM1 is still limited.STMN1 is a microtubule-binding protein which can inhibit the assembly of microtubule dimer or promote depolymerization of microtubules.It was reported as a major responsive factor of paclitaxel resistance for clinical chemotherapy of tumor patients.But the function of abnormally high level of STMN1 and its regulation mechanism in cancer cells remain unclear.In this study,we used public database and tissue microarrays to analyze the expression pattern of FoxM1 and STMN1 and found a strong positive correlation between FoxM1 and STMN1 in multiple types of cancer.Lentivirus-mediated FoxM1/STMN1-knockdown cell lines were established to study the function of FoxM1/STMN1 by performing cell viability assay,plate clone formation assay,soft agar assay in vitro and xenograft mouse model in vivo.Our results showed that FoxMl promotes cell proliferation by upregulating STMN1.Further ChIP assay showed that FoxM1 upregulates STMN1 in a transcriptional level.Prognostic analysis showed that a high level of FoxM1 and STMN1 is related to poor prognosis in solid tumors.Moreover,a high co-expression of FoxM1 and STMN1 has a more significant correlation with poor prognosis.Our findings suggest that a general FoxMl-STMN1 axis contributes to cell proliferation and tumorigenesis in hepatocellular carcinoma,gastric cancer and colorectal cancer.The combination of FoxM1 and STMN1 can be a more precise biomarker for prognostic prediction.