Takeshi Uchitane

A Tool for Parameter-space Explorations☆

Abstract A software for managing simulation jobs and results, named “OACIS”, is presented. It controls a large number of simulation jobs executed in various remote servers, keeps these results in an organized way, and manages the analyses on these results. The software has a web browser front end, and users can submit various jobs to appropriate remote hosts from a web browser easily. After these jobs are finished, all the result files are automatically downloaded from the computational hosts and stored in a traceable way together with the logs of the date, host, and elapsed time of the jobs. Some visualization functions are also provided so that users can easily grasp the overview of the results distributed in a high-dimensional parameter space. Thus, OACIS is especially beneficial for the complex simulation models having many parameters for which a lot of parameter searches are required. By using API of OACIS, it is easy to write a code that automates parameter selection depending on the previous simulation results. A few examples of the automated parameter selection are also demonstrated.

Traffic Simulation of Kobe-City

A traffic simulation of Kobe-city was carried out. In order to simulate an actual traffic flow, a road network was constructed utilizing a high-quality digital map data, and an origin-destination information of vehicles was estimated by a geographical population distribution data. The result obtained in this way was incompatible with the traffic census data due to the differences between the simulation and actual traffic, such as routing, OD information and so on. In order to improve the reproducibility of the traffic flow, the parameter search whose adjustable parameter was the speed limit of the road was conducted. This adjustment showed that reproducibility improves. Further improvement of the reproducibility needs to reconsideration of the routing algorithm.

Experimental study for multi-objective PSO with single objective guide selection

Multi-objective particle swarm optimization has two different points from single objective one. The first point is guide position selection methods for personal best and global best. The second one is the usage of an archive to preserve good positions for Pareto optimal set. In this paper, we consider a guide selection problem in multiobjective particle swarm optimization. A selection method for the personal best that depends on one objective function among plural objective functions is presented. Then, a selection method for the global best that selects among the archived position due to one objective function is presented. The performances of the proposed methods are evaluated by the benchmark problems for the evolutionary multiobjective optimization algorithms.

An open-source job management framework for parameter-space exploration: OACIS

We present an open-source software framework for parameter-space exploration, named OACIS, which is useful to manage vast amount of simulation jobs and results in a systematic way. Recent development of high-performance computers enabled us to explore parameter spaces comprehensively, however, in such cases, manual management of the workflow is practically impossible. OACIS is developed aiming at reducing the cost of these repetitive tasks when conducting simulations by automating job submissions and data management. In this article, an overview of OACIS as well as a getting started guide are presented.

Applying evolution strategies for biped locomotion learning in RoboCup 3D Soccer Simulation

This paper addresses parameter tuning methods for bipedal locomotion of a humanoid model in the RoboCup 3D Soccer Simulation environment. A gait pattern of this humanoid is generated by a desired foot trajectory, joint control systems and nonlinear oscillators. To build a good gait pattern, the parameters of the walking system should be adjusted suitably. In this paper, a usage of evolution strategies that is depending on only a performance evaluation of the robot, is considered for adjusting the parameters. We apply two type evolution strategies in order to tune the parameters. The one is an evolution strategy with mask operation where the portion of individual to avoid mutation. The other is a covariance matrix adaptation evolution strategy. Numerical simulation studies are carried out to evaluate the performance of the proposed approaches by using the RoboCup 3D Soccer Simulator.

An Environment for Combinatorial Experiments in a Multi-agent Simulation for Disaster Response

We present a research environment for combinatorial experiments for the RoboCupRescue Simulation, which is a platform for the study of disaster-relief strategies using multi-agent simulations. To simulate the agents in disaster-relief situations in the RoboCupRescue Simulation, it is necessary to implement a wide variety of algorithms for tasks such as such as group formation, path planning, and task allocation. Recently, we proposed a modular framework, the Agent Development Framework, that enables researchers to implement, study, and test each algorithm independently. Because the algorithms developed in this framework are mutually replaceable, it is possible to combine algorithms developed by different researchers. In this study, we further propose an experimental environment to efficiently handle the experiments of a huge number of possible combinations of the algorithms. As a demonstration, we test various combinations of the algorithms developed by the participants of RoboCup 2017 and show that there indeed exists a set of the algorithms that is superior to the original ones developed by each team.

A General Swarm Intelligence Model for Continuous Function Optimization.

We consider a general form of the swarm intelligence as a function optimization tool. This form is derived from a basis of mathematical swarming differential equation model, where several parameters are included in the model. These parameters are corresponding to a repulsion effect, an attractive effect and a gradient direction. We mainly consider a repulsion effect and unknown gradient estimation in this study. The nature of the proposed model by some typical numerical simulation results is described. Then, the numerous simulation results show that the behaviors of the swarm will change significantly, for example, aggregation and clustering by parameter setting. We are able to see basic behaviors of the swarm intelligence by the introduced model, the model could give us the insight to understand search behavior of swarm intelligence.

Proposed Environment to Support Development and Experiment in RoboCupRescue Simulation

The RoboCupRescue Simulation project is a test bed for multi-agent systems research for disaster relief. However, researchers have to implement many types of algorithm and require a complicated procedure for experiments, which places a heavy burden on them. Therefore, we propose an environment that integrates an agent-development framework and an experiment-management system to support researchers.

A study on multi-objective particle swarm model by personal archives with regular graph

Multi-objective evolutionary optimization algorithms have been received much attention in recent years, since a set of Pareto optimal candidate is provided by a single run. Generally, it is required that the provided candidates of Pareto solutions cover the Pareto front widely and uniformly. To achieve this requirement, there has been proposed a lot of variants of multi-objective evolutionary algorithms including multi-objective particle swarm models. We are able to see two major differences in the previously proposed multi-objective particle swarm models, the one is a use of single external archive and depending on additional random effect to maintain particle diversity in the swarm. In this paper, we propose more natural way to apply multi-objective optimization of particle swarm, where we introduce a personal archive that stores non-dominated candidates in each particle history. By numerical examples, the proposed method is able to provide better Pareto candidates without an additional random effect on the swarm model.