Toshiyuki Yasuda

Accelerating steady-state genetic algorithms based on CUDA architecture

Parallel processing using graphic processing units (GPUs) have attracted much research interest in recent years. Parallel computation can be applied to genetic algorithms (GAs) in terms of the processes of individuals in a population. This paper describes the implementation of GAs in the compute unified device architecture (CUDA) environment. CUDA is a general-purpose computation environment for GPUs. The major characteristic of this study is that a steady-state GA is implemented on a GPU based on concurrent kernel execution. The proposed implementation is evaluated through four test functions; we find that the proposed implementation method is 3.0–6.0 times faster than the corresponding CPU implementation.

A homogeneous mobile robot team that is fault-tolerant

Abstract This paper introduces a design methodology of a fault-tolerant autonomous multi-robot system (MRS). An important fundamental topic for this type of system is the design of an on-line autonomous behavior acquisition mechanism that is capable of developing cooperative roles as well as assigning them to a robot appropriately in a noisy embedded environment. Our approach is to apply reinforcement learning that adopts the Bayesian discrimination method for segmenting a continuous state space and a continuous action space simultaneously. In addition, a neural network is provided for predicting the average of the other robots’ postures at the next time step in order to stabilize the reinforcement-learning environment. Computer simulations are conducted to illustrate the fault-tolerance of our MRS against a system change that occurs after the MRS achieves stable behavior.

MBEANN: mutation-based evolving artificial neural networks

A novel approach to topology and weight evolving artificial neural networks (TWEANNs) is presented. Compared with previous TWEANNs, this method has two major characteristics. First, a set of genetic operations may be designed without recombination because it often generates an offspring whose fitness value is considerably worse than its parents. Instead, two topological mutations whose effect on fitness value is assumed to be nearly neutral are provided in the genetic operations set. Second, a new encoding technique is introduced to define a string as a set of substrings called operons. To examine our approach, computer simulations were conducted using the standard reinforcement learning problem known as the double pole balancing without velocity information. The results obtained were compared with NEAT results, which is recognised as one of the most powerful techniques in TWEANNs. It was found that our proposed approach yields competitive results, especially when the problem is difficult.

A self adaptive hybrid enhanced artificial bee colony algorithm for continuous optimization problems

The artificial bee colony (ABC) algorithm is one of popular swarm intelligence algorithms that inspired by the foraging behavior of honeybee colonies. To improve the convergence ability, search speed of finding the best solution and control the balance between exploration and exploitation using this approach, we propose a self adaptive hybrid enhanced ABC algorithm in this paper. To evaluate the performance of standard ABC, best-so-far ABC (BsfABC), incremental ABC (IABC), and the proposed ABC algorithms, we implemented numerical optimization problems based on the IEEE Congress on Evolutionary Computation (CEC) 2014 test suite. Our experimental results show the comparative performance of standard ABC, BsfABC, IABC, and the proposed ABC algorithms. According to the results, we conclude that the proposed ABC algorithm is competitive to those state-of-the-art modified ABC algorithms such as BsfABC and IABC algorithms based on the benchmark problems defined by CEC 2014 test suite with dimension sizes of 10, 30, and 50, respectively.

An incremental approach to an evolutionary robotic swarm

Swarm robotics is the research field of multi-robot systems which consist of many homogeneous autonomous robots without any type of global controllers. In this paper, an evolutionary robotics approach, i.e., the method that the robot controllers are designed by evolutionary algorithms with artificial neural networks, is applied. And we apply an incremental approach to evolution within the context of evolutionary robotics, so as to obtain controllers that has high generalization capability. As a benchmark of robotic swarm, cooperative food-foraging problems are conducted to examine their performance. In particular, The flexibility and scalability of the neural controllers are discussed.

An evolutionary approach to sudoku puzzles with filtered mutations

Sudoku puzzles are logical number placement puzzle games. They are classified as combinatorial optimization problems and are NP-complete. To solve problems in this complexity class, metaheuristic approaches, such as genetic algorithms (GAs), are often adopted. However, conventional GAs with random swap mutations suffer from slow convergence, especially in extremely difficult sudoku puzzles, in which only a few given numbers are provided. This paper proposes a GA with sophisticated genetic mutations that mitigate the worsening of fitness values. The comparisons between the conventional method and the proposed method are conducted mainly from the viewpoints of success rate.

Coordinating the adaptive behavior for swarm robotic systems by using topology and weight evolving artificial neural networks

Swarm robotics (SR) is the research field of multirobot systems, which consist of many homogeneous autonomous robots without any types of global controllers. Generally, since a task given to this system cannot be achieved by a single robot, cooperative behavior is expected to emerge in a robotic swarm by a certain mechanism, which is through the interactions among robots or with an environment. In this paper, an evolutionary robotics approach, in which robot controllers are designed by evolving artificial neural networks, is adopted. Among the many approaches to evolving artificial neural networks, two approaches, NEAT and MBEANN are adopted for conducting computer simulations. Although a conventional neural network has a fixed topology and evolves only with its synaptic weights, NEAT and MBEANN evolve not only with their synaptic weights, but also with their topologies. As a benchmark for swarm robotics, cooperative package-pushing problems using ten autonomous robots are conducted to evaluate their performance. The behavioral characteristics that emerge are then discussed.

A Reinforcement Learning Technique with an Adaptive Action Generator for a Multi-robot System

We have developed a new reinforcement learning (RL) technique called Bayesian-discrimination-function-based reinforcement learning (BRL). BRL is unique, in that it does not have state and action spaces designed by a human designer, but adaptively segments them through the learning process. Compared to other standard RL algorithms, BRL has been proven to be more effective in handling problems encountered by multi-robot systems (MRS), which operate in a learning environment that is naturally dynamic. Furthermore, we have developed an extended form of BRL in order to improve the learning efficiency. Instead of generating a random action when a robot functioning within the framework of the standard BRL encounters an unknown situation, the extended BRL generates an action determined by linear interpolation among the rules that have high similarity to the current sensory input. In this study, we investigate the robustness of the extended BRL through further experiments. In both physical experiments and computer simulations, the extended BRL shows higher robustness and relearning ability against an environmental change as compared to the standard BRL.

Evolving robot controllers for a homogeneous robotic swarm

A homogeneous robotic swarm has a great potential to achieve a highly complex task that cannot be solved by a single robot. However, it would not be so easy to develop an appropriate collective behavior because of the intrinsic large redundancy in itself. In this paper, a robot controller is designed with an evolving artificial recurrent neural network (EANN) to develop coordinated collective behavior autonomously. However, it is well-known that an EANN shows the performance strongly dependent on its topological structure. In addition to this, no general guidelines have been known to the question of how to design an appropriate topological structure of an EANN. Therefore, as the beginning of solving this question, an approach to giving the evolvability of the topological structure at the limited part between the hidden layer and the output layer is introduced. A series of computer simulations are conducted to draw a conclusion.

Self-organized flocking of a mobile robot swarm by topological distance-based interactions

Self-organized flocking of robotic swarms has been investigated for approximately twenty years. Most studies are based on a computer animation model named Boid. The Boid model reproduces flocking motion by three simple behavioral rules: collision avoidance, velocity matching, and flock centering. Boid-like flocking motions are generated from the interacting positions and orientations of neighboring robots. This paper investigates a technique by which interactions among neighboring robots in a swarm are determined by their topological separations. The effectiveness of the proposed method is quantitatively evaluated in real robot experiments.