Over the past few years, deep neural networks (DNNs) have shown outstanding performance in a wide range of domains. However, DNNs have been found to be vulnerable to adversarial examples (AE). AE are inputs that are designed to cause poor performance to a predictive machine learning model. Adversarial attacks are classified into two categories: targeted attacks and non-targeted attacks. In a targeted attack on multi-class classifiers, there will be multiple AEs that mislead models to a class other than the true class. As one of the black-box attacks on computer vision, a method of generating adversarial examples using Differential Evolution (DE) has been reported. This attack method named one pixel attack is very effective because the output of the model can be greatly changed by modifying a few pixels of the input image. However, in order to acquire multiple AEs in a targeted attack, it is necessary to repeatedly execute the targeted attack while changing the target class. In this case, multiple trials are required, and the number of accesses to the model increases in proportion to the number of classes. Therefore, we propose a new method to acquire multiple AEs with a single run of one pixel attack. In the proposed method, the objective function in a non-targeted attack is regarded as a multimodal landscape with multiple solutions. Then, a penalty is dynamically added to the objective function in this multimodal function to search for multiple solutions in order. Additionally, to improve the search efficiency of one pixel attack, Rank-based DE (RDE), which is an improved method of DE, is introduced. We conducted experiments using some typical machine learning models and showed that multiple AEs can be efficiently acquired by the proposed method.

On December 16, 2022, Open Space Discussion 2022 (OSD2022) was held as a pre-event of the 2022 Symposium on Evolutionary Computation. This event was motivated to provide an opportunity to share, discuss, and create future directions in evolutionary computation. This paper provides an event report for OSD2022, including a summary of the discussions made as well as participant feedback.

In recent years, digital twin technologies, which reproduce various data from the real world in virtual space, have been attracting attention due to a rapid increase in computational resources. In the field of crowd control, multiagent simulation (MAS) enables precise simulation of crowd movements and is anticipated to be used for evaluating how to guide crowds, such as evacuation efficiency and potential risks of crowd accidents. In this context, the fidelity of the crowd simulation is a crucial factor. In the Evolutionary Computing Competition 2022, an optimization competition was held to estimate pedestrian traffic information from observed data using MAS. This paper outlines the formulation of the optimization problem and discusses the problem’s characteristics by introducing the optimization algorithms and their results provided by the participating teams.

In this paper, we study a distributed Q-learning approach for multi-objective optimization of epistatic binary problems and investigate its performance. The Q-learning based method assigns an agent per objective function, specifies a state as a solution and a position where an agent can act and defines an action as a 1-bit mutation operator. The method also introduces conditional state transitions, where an agent moves to a new state only if the chosen action improves the best solution found in the current episode. Otherwise, the agent discards previously chosen actions and continues sampling from the same state until a threshold is reached. We investigate the proposed method solving MNK-landscapes with 100 bits varying the number of objectives from 2 to 4 and the number of epistatic interactions from 1 to 20. We also compare with a multi-objective random bit climber moRBC that also implements a 1-bit neighborhood search and NSGA-II and MOEA/D, two well known multi-objective algorithms representatives of Pareto dominance and decomposition based approaches, to better understand the algorithm’s search behavior and performance on epistactic problems of increased difficulty.

This paper proposes an evolutionary algorithm named MFEA/OS (Multi-Factorial Evolutionary Algorithm based on Objective Similarity) for multi-factorial optimization that aims to optimize multiple objective functions simultaneously. In evolutionary multi-factorial optimization, each solution in the population is associated with an objective function to be optimized. Evolutionary variation such as crossover is applied to two solutions, even if they are associated with different objective functions. This interaction is expected to enhance the simultaneous optimization of multiple objective functions. However, generating new offspring solutions from parent solutions associated with dissimilar objective functions would actually be harmful to their simultaneous optimization. For each pair of objective functions, the proposed MFEA/OS calculates the difference in the distributions of their associated solutions in the variable space as the objective similarity. The proposed MFEA/OS then encourages evolutionary variation between two solutions associated with similar objective functions. To verify the effects of the proposed MFEA/OS, this work uses continuous and discrete test problems that can adjust the similarities among multiple objective functions. Experimental results show that the proposed MFEA/OS can estimate the objective similarities in continuous and discrete test problems and achieves higher multi-factorial optimization performance than conventional algorithms in both test problems with correlated objective functions.

AεSεH is one of the evolutionary algorithms used for many-objective optimization. It uses ε-dominance during survival selection to sample from a large set of non-dominated solutions to reduce it to the required population size. The sampling mechanism works to suggest a subset of well distributed solutions, which boost the performance of the algorithm in many-objective problems compared to Pareto dominance based multi-objective algorithms. However, the sampling mechanism does not select exactly the target number of individuals given by the population size and includes a random selection component when the size of the sample needs to be adjusted. In this work, we propose a more elaborated method also based on ε-dominance to reduce randomness and obtain a better distributed sample in objective-space to further improve the performance of the algorithm. We use binary MNK-landscapes to study the proposed method and show that it significantly increases the performance of the algorithm on non-linear problems as we increase the dimensionality of the objective space and decision space.