Generative AI has become a cornerstone of modern AI research, transforming approaches to image generation and language processing through technologies like Large Language Models (LLMs). Tools such as ChatGPT and Stable Diffusion showcase the creative capabilities of AI and its profound societal impact. Simultaneously, stochastic optimization methods, including evolutionary computation, remain indispensable for solving complex challenges. The growing interplay between generative AI and optimization techniques underscores their complementary strengths. This paper surveys recent developments in leveraging the synergy between generative AI and stochastic optimization, shedding light on emerging trends and future directions.

This paper describes a method for estimating the ripeness of persimmons from images. When a human visually judges the CC value, which is a numerical value indicating the degree of ripeness shown on a color chart for fruit, the judged value is not always the same for everyone, and there is a problem that it is not quantitative. We propose a method for measuring persimmon fruit peel color using a spectrophotometer and calculating the CC value. We also propose an image processing method that takes into account the characteristics of fruit peel color variation as a data augmentation method for training deep learning models. In the second half of the paper, we show the possibility of investigating the relationship between color and ripeness through experiments in which a wider range of CC values than those in the data set are trained. The follow-up study of CC values for each fruit also indicated the possibility of approximately predicting the optimal harvest time for each fruit.

There is a lot of previous research on cultivation management based on mathematical engineering. In some previous research, the expected profit of cost constrained cultivation management for multiple agricultural fields with sensors is maximized under the condition that growing states are unknown. But in the previous research, it is assumed that the same crop will be grown in all agricultural fields and crop prices do not fluctuate. In this research, cost constrained cultivation management with sensors considering fluctuation of crop price under the condition that different crops can be grown in the agricultural fields and growing states are unknown is studied. The cultivation management problem is modeled by Markov decision processes as previous research. The expected profit is maximized with reference to a Bayes criterion using dynamic programming in proposed method. The effectiveness of the proposed method is shown by some computational examples. Adaptive cost adjustment between the different crops is confirmed in the computational examples.

In this paper, we consider an effective search method for large scale combinatorial optimization problems, only by means of neighborhood operations, not by means of such operation as crossover in genetic algorithm. The fundamental ideas on which our method is based are the following: 1) Many different neighborhood operations, which consist of the iterations of unit neighborhood operations, are applied to solutions. 2) The probability distribution of the objective function values of the neighborhood solutions is estimated, from the data obtained in the search process. 3) The neighborhood operation, which maximizes the expected value of the amount of the improvement of the current solution, is selected to be applied. From these ideas, a new method for searching for solutions is constructed, on the basis of the self-convolution and the inverse self-convolution. We have applied the local search method, the previously proposed method, and the newly proposed method to traveling salesman problems and maximum satisfiability problems. The effectiveness of the newly proposed method is shown by the computational experiments.

The authors aim to develop an integrated interface, Bio Discovery OS (BioDOS), that can present DNA sequences encoding various biomolecular networks corresponding to a desired behavior. Its development is being conducted as part of a team-based research project (Core Research for Evolutional Science and Technology; CREST). This CREST is broadly divided into the design of sequences by hybrid AI to generate a large number of candidate biomolecular networks followed by their refinement (Dry study), and the construction of a system based on isomorphic combinations guaranteed by mathematical models and biological experiments to confirm its operation (Wet study). Especially, the authors focus on determining the similarity with existing networks in the Dry study. It is required to narrow down the large number of candidate networks generated by machine learning and reinforcement learning by matching the successes of existing articles, biological experiments, and numerical simulations. In this paper, the authors aim to propose a method for discovering useful articles for BioDOS using various machine learning methods such as deep learning, and construct an article recommendation system for BioDOS using the proposed method. Numerical experiments are conducted to verify the effectiveness of the proposed method and the constructed recommendation system.

In this research, we propose a novel method to accurately predict snow cover rate on photovoltaic panels in snowy areas for precise photovoltaic power generation predicting. Our proposed method utilizes weather data and past snow cover rates as input into a deep learning model composed of an autoencoder for compressing weather data dimensions and a bidirectional LSTM network for time series modeling. The deep learning model can predict hourly snow cover rates during daylight hours for the target prediction period. In experiments, we compare the prediction accuracy against using various data patterns and various deep learning models. As a result of the evaluation, the proposed method achieved the highest accuracy. Moreover, applying the predicted snow cover rates to photovoltaic power predicting substantively improves accuracy compared to using only weather data. These results demonstrate the effectiveness of the proposed method.

The role of data centers has been becoming increasingly crucial in recent years, because there has been a growing emphasis on decentralizing data processing in IoT networking. The ability to swiftly respond to incoming requests has emerged as a pivotal metric for assessing data center systems. Load balancing challenges in data centers have persistently posed hurdles, affecting performance enhancement, availability improvement, and cost reduction. In this paper, we delve into the load balancing challenges in SDN-enabled data centers and introduce two novel load balancing methods. One method is the weighted minimum response time approach, while the other employs learning automata for load balancing. The former tackles the issue of server load fluctuations without the bias found in static weighting methods, whereas the latter seeks to equalize server loads for queries with established access patterns. We conducted straightforward computer simulations to evaluate the proposed methods. The results of computer experiments indicate the effectiveness of the proposed methods.

Deep reinforcement learning is a machine learning method that combines deep learning and reinforcement learning. Deep Q-network (DQN) is one of the typical methods of deep reinforcement learning. DQN uses Convolutional Neural Network (CNN) which can extract features from the input images. We have applied DQN method to the mobile robot navigation problem. The values of hyper-parameters, including the network structure of DQN, and the reward function used in the DQN algorithm, have been determined empirically. In this study, we attempt to optimize both of the values of hyper-parameters and reward function of deep reinforcement learning by using Bayesian optimization. We realized to optimize the values of hyper-parameters including the network structure of DQN, and the reward function by using Optuna, a framework of Bayesian optimization. We confirmed that the values of hyper-parameters and reward function obtained by Optuna have higher learning performance than that by empirical method.

Facial Expression Recognition has been studied for many years; however, it remains a challenging task in real-world environments due to complex backgrounds, varying illumination conditions, and online processing issues. In this study, we propose a deep learning model, CAER-Net-RS, by leveraging multiple training datasets. The proposed model integrates three neural networks: the Face Network, the Context Network, and the Adaptive Network. Different datasets are employed for the pretraining of these networks: the facial expression image dataset RAF-DB for the Face Network, the scene image dataset Places365-Standard for the Context Network, and the CAER-S dataset for the Adaptive Network. In the experiment, the proposed model achieved an average recognition accuracy of 85.20% across seven types of facial expressions, compared to 70.92% for the conventional Context-Aware Emotion Recognition Network (CAER-Net).

To improve the accuracy of dam inflow forecasting, we propose a prediction model using fine-tuning with sLoRA (self-Low Rank Adaptation), a novel technique that applies the lightweight fine-tuning method of Low Rank Adaptation (LoRA) to itself. This model utilizes an MLP(Multi Layer Perceptron) as the base model and performs fine-tuning on the neighborhood set extracted by k-NN(k Nearest Neighbor) for each query. Experimental results using two dam datasets demonstrate that the proposed method outperforms both the MLP-only and k-NN-based models in terms of accuracy. Furthermore, the prediction accuracies, measured by both RSR and NSE indexes, are categorized as Very Good.

An antenna beamforming system with feedback phase control is proposed, and an analysis on the control stability is described. When the number of the antenna elements is two, the phases are controlled depending on the residual phase difference of the signals. It can be realized with a simple feedback system. When the number of the antenna elements is more than two, there are plural phase differences, and plural control loops are required. It makes the stability analysis complex. In this paper, a control system with small number of feedback loops is proposed, and an analysis method with eigen values of the transfer matrix is described.

Optical fiber composite overhead ground wire (OPGW), which is used as communication lines by electric power companies, sometimes freezes and causes communication faults in winter season because of rainwater immersion into aluminum pipes which envelop optical fibers (referred to as OP unit). The rainwater in OP unit causes corrosion reaction with aluminum, and generates hydrogen gas. If the hydrogen gas penetrates the core of an optical fiber, optical loss at the wavelength of 1.24 and 1.625 µm increases. The author proposes an estimation method of water immersion amount in OP unit, based on continuous measurement data of an optical time domain reflectometer (OTDR) using 1.24/1.625 µm.

Pain is an indicator of various diseases and injuries, and research has been conducted to objectively evaluate this pain using magnetoencephalography, fMRI, and electroencephalography (EEG). This paper focuses particularly on EEG, which can be evaluated simply. Previous studies on pain have simulated brief stimulation with heat or pulsed electric current, but in this research, sustained pain was inflicted on research subjects using foot pressure point stimulation, which is used for health purposes. The measured data of the somatosensory cortex EEG at that time was analyzed by RMS of the waveform and Hjorth parameters used in real-time signal frequency analysis. The results of the linear mixed-effects model showed that the Activity values of the RMS and Hjorth parameters showed significant regression with stimulus intensity and subjective evaluation, indicating the possibility of objective evaluation of persistent pain using EEG. In addition, the Mobility and Complexity of Hjorth parameters showed no relationship with stimulus intensity, suggesting that the change in EEG frequency due to pain may be non-stationary. Finally, the temporal coefficient of variation of RMS was smaller than that of Activity, suggesting that RMS provides the most stable objective evaluation.

Handicapped people who can only operate one pushrim of a self-operated wheelchair, such as hemiplegic patient and upper extremity amputees, are physically burdened by the need to correct the driving direction with their feet. This study proposes a control system of power-assisted wheelchairs to realize the desired driving in arbitrary direction with only one pushrim operation. The proposed system measures forearm EMG signals when the user grasps the pushrim while flexing or extending the wrist in accordance with the driving direction and estimates the target driving position by fuzzy inference using EMG signals and pushing torque. In addition, an operation training system to improve the EMG and pushrim operation proficiency is also introduced. Some detailed experimental verification including practical driving demonstrates the effectiveness of the proposed control system.

A design of sliding mode synchronize controller based on model reference control scheme and the experimental results of wheel chair bi-wheel running test are described. The controller with manual wheel chair reference model that estimated wheel drive force as the other wheel dive one is input performs one-hand driven wheel chair. The robust bi-wheel synchronization under the appropriate drive force feedback gain demonstrates one-hand driven wheel chair not only rectilinear motion but also clock wise (or counter clock wise) rotation and pivot turn.

Iterative Learning Control (ILC) is a method to achieve tracking a specific reference by repeatedly updating control inputs. The learning algorithm of the ILC is conducted by using previous error information and learning parameters for systems that perform the same operation repeatedly. The ILC design requires an impulse response model for the learning coefficient matrix. However, it suffers from estimating an impulse response model in case where the measurement noise is present. Therefore, the study proposes a method for estimating impulse response to mitigate the influence of measurement noise. The proposed method estimates an impulse response projected to the linear space generated by the nominal model. The effectiveness of the proposed method is illustrated through numerical simulation results.