Kazutoshi Sakakibara

Multi-objective approaches in a single-objective optimization environment

This paper presents two new approaches for transforming a single-objective problem into a multi-objective problem. These approaches add new objectives to a problem to make it multi-objective and use a multi-objective optimization approach to solve the newly defined problem. The first approach is based on relaxation of the constraints of the problem and the other is based on the addition of noise to the objective value or decision variable. Intuitively, these approaches provide more freedom to explore and a reduced likelihood of becoming trapped in local optima. We investigated the characteristics and effectiveness of the proposed approaches by comparing the performance on single-objective problems and multi-objective versions of those same problems. Through numerical examples, we showed that the multi-objective versions produced by relaxing constraints can provide good results and that using the addition of noise can obtain better solutions when the function is multimodal and separable.

Prediction of the O -glycosylation sites in protein by layered neural networks and support vector machines

O-glycosylation is one of the main types of the mammalian protein glycosylation, which is serine or threonine specific, though any consensus sequence is still unknown. In this paper, a layered neural network and a support vector machine are used for the prediction of O-glycosylation sites. Three types of encoding for a protein sequence within a fixed size window are used as the input to the network, that is, a sparse coding which distinguishes all 20 amino acid residues, 5-letter coding and hydropathy coding. In the neural network, one output unit gives the prediction whether a particular site of serine or threonine is glycosylated, while SVM classifies into the 2 classes. The performance is evaluated by the Matthews correlation coefficient. The preliminary results on the neural network show the better performance of the sparse and 5-letter codings compared with the hydropathy coding, while the improvement according to the window size is shown to be limited to a certain extent by SVM.

Improvements of the Traffic Signal Control by Complex-Valued Hopfield Networks

The phase synchronization in the complex-valued Hopfield network has been shown to be effective for a signal control in an area-wide urban traffic flow control. The basic idea of the original method is to attain the global effectiveness as a weighted summation of the local effectiveness. And the complex-valued Hopfield network is designed to converge to such an optimal state through the appropriate interaction between the neurons which model the traffic signals. As the result, the network possesses the energy function which expresses the global effectiveness, whose leading term is given by the summation of the substantial traffic flows under the given offset. Thus, it is a bottom-up approach to optimize the global effectiveness as the total of the local effectiveness. In this paper, two different approaches are introduced, and added to or compared with the above approach. The first approach is the feedback from the real time information of local traffics. The purpose of the feedback is to decrease the differences of the disadvantages among conflicting flows, which are measured by a congestion or the number of waiting vehicles. The addition of the feedback to the original method shows that the local feedback works as a pinpoint control on a local congestion, while keeping the total effectiveness especially in regular traffic patterns. The second is a top-down approach to attain the global optimization by real-coded genetic algorithms. The proposed GA directly searches the effective offset using a traffic simulator to calculate the average traveling time for the evaluation. Therefore, genetic operations are designed for a small size population and a real-code in a torus space. The best offsets obtained by GA reduce the average traveling time by 2%~7% compared with the results obtained by the original approach.

A multiobjectivization approach for vehicle routing problems

This paper presents a new approach for vehicle routing problems (VRPs), which are defined as problems of minimizing the total travel distance. The proposed approach treats VRPs asmulti-objective problems using the concept of multiobjectivization. The multiobjectivization approach translates single-objective optimization problems into multi-objective optimization problems and then applies EMO to the translated problem. In the proposed approach, a newly defined objective related to assignment of customers is added, because the assignment has a more important influence on the search results than routing in VRPs.We investigated the characteristics and effectiveness of the proposed approaches by comparing the performance on conventional approaches and the proposed approaches.

Effective Integration of Imitation Learning and Reinforcement Learning by Generating Internal Reward

This paper describes an integrative machine learning architecture of imitation learning and reinforcement learning. The learning architecture aims to help integration of the two learning process by generating internal rewards. After observing superiors, human learners usually start practicing through trial and error. Humans usually learn tasks through both imitation learning and reinforcement learning. Imitation learning and reinforcement learning should be harmonized as an effective and integrative learning system. A simple reinforcement learning requires a huge amount of trials and errors in an agents learning phase. However, imitation learning can reduce the amount of time. Based on this idea, the composition of reinforcement learning and imitation learning is proposed as an integrative machine learning architecture. In this paper, an additional internal reward system, which is generated by the learner agent, is introduced to achieve this goal. The learning architecture is evaluated through an experiment and the effectiveness of the integration is examined.

PHASE DYNAMICS OF COMPLEX-VALUED NEURAL NETWORKS AND ITS APPLICATION TO TRAFFIC SIGNAL CONTROL

Complex-valued Hopfield networks which possess the energy function are analyzed. The dynamics of the network with certain forms of an activation function is de-composable into the dynamics of the amplitude and phase of each neuron. Then the phase dynamics is described as a coupled system of phase oscillators with a pair-wise sinusoidal interaction. Therefore its phase synchronization mechanism is useful for the area-wide offset control of the traffic signals. The computer simulations show the effectiveness under the various traffic conditions.

Analysis on battery storage utilization in decentralized solar energy networks based on a mathematical programming model

We study a decentralized solar energy network composed of multiple clusters, which correspond to households equipped with PV units, rechargeable batteries, electrical appliances, and an electric power router. Decentralized solar energy network is a new grid systems toward independence from existing power grid, and solar energy is main energy source for the decentralized energy network. Each cluster has a battery storage to use the renewable energy effectively, where battery degradation cannot be overlooked for long-term, persistent operation. This paper proposes an optimal power distribution minimizing the battery degradation on decentralized energy network. Because the battery degradation is unavoidable phenomenon for the battery utilization, we solve the optimal power distribution to keep the degradation minimum. The proposed approach limits the charge/discharge speeds and cycles at mixed integer programming formulation and achieves the optimal utilization of the renewable energy with less charge/discharge cycles. Experimental results using the real measured data of the power generation and consumption show the connection between the parameters for the battery degradation and the system performance.

Genetics-Based Machine Learning Approach for Rule Acquisition in an AGV Transportation System

We propose an autonomous decentralized method for multiple AGV robots under uncertain delivery requests. Transportation route plans of AGV robots are expected to minimize the transportation time without collisions among the robots in the systems. In our proposed methods, each robot as an agent computes its transportation route by referring to the static path information, and it exchanges its route plan each other. Once collisions are detected, one of the two agents chosen by a negotiation rule modifies its route plan. The rule consists of a condition-part and an action-part, and one rule which matches to the conditions of two agents under negotiation is selected from a set of rules. The rules are generated and improved by a genetic based machine learning approach, where a set of rules is represented symbolically as an individual of genetic algorithms, and fitness of each individual is determined according to the total travel time of the AGVs and the adequacy of the condition-parts of the rules.

Autonomous distributed approaches for pickup and delivery problems with time windows

We consider the pickup and delivery problem with time windows as one of the practical transportation problems. The problem requires that any paired pickup and delivery locations have to be served by one vehicle and the pickup location has to be scheduled before the corresponding delivery location in the route. In this paper, to search a set of routes close to the optimal one, we propose autonomous distributed approaches based on the search space decomposition for the problem. In this approach, first, the search space is divided into sub-spaces based on the number of customers loaded on each vehicle. Then, the genetic algorithm is applied to these sub-spaces. The effectiveness of the proposed approach is evaluated by computational experiments.

Prediction of the O-Glycosylation with Secondary Structure Information by Support Vector Machines

Mucin-type O-glycosylation is one of the main types of the mammalian protein glycosylation. It is serine (Ser) or threonine (Thr) specific, though any consensus sequence is still unknown. In this report, support vector machines (SVM) are used for the prediction of O-glycosylation for each Ser or Thr site in the protein sequences. 29 mammalian protein sequences are selected from UniProt8.0, and its structure information is obtained from Protein Data Bank (PDB). A protein subsequence with a prediction target of Ser or Thr site at the center is used as input to SVM, and its amino acid sequence information, and the secondary structure or accessibility, which are calculated by DSSP from PDB data, are encoded as an input data. The results of the preliminary experiments show the effectiveness of the local structure information added to the sequence information.