Kenji Iwadate

An Artificial Neural Network Based on the Architecture of the Cerebellum for Behavior Learning

In the last decade, artificial intelligence (AI) pervades every aspect of our lives. However, there is a gap between AI-based machine behavior and human in natural communication. The behavior of most AI is determined as a task list generated by engineers, but to obtain high-level intelligence, AI needs the ability to cluster tasks from circumstances and learn a strategy for achieving each task. In this study, we focus on the human brain architecture that gives it the ability to self-organize and generalize sensory information. We propose an Artificial Neural Network (ANN) model based on that architecture. We describe a cerebellum-based ANN model (C-ANN) and verify its capacity to learn from the phototaxic behavior acquisition of a simple two-wheeled robot. As a result, the controller of the robot is self-organized to be simple and able to achieve positive phototaxis. This result suggests that the proposed C-ANN model has the capability of supervised learning.

Behavior emergence of a virtual creature living in complex environments

This study aimed at establishing a new computer-aided animation method using agent-based and physics modeling-based animation. The specific problem we addressed was to install adaptive behavior in a virtual creature placed in a complex environment, and to create its animated behavior automatically. The virtual creature is regarded as an autonomous agent who has sensors, actuators, and controllers. An artificial neural netword (ANN) and a central pattern generator (CPG) were adopted as the controllers. An optimization algorithm was introduced to train the controllers. Numerical experiments proved that the virtual creature acquires effective motions (walking, swimming) to move toward its destination, and to avoid obstacles and other creatures.

Realization of dynamic interaction of virtual creatures with neural controller

The objective of this study is to establish a new computer-aided animation method using agent-based and physics-modeling-based animation. The specific problem considered in this paper is the realization of the adaptive behavior of a virtual creature placed in a complex environment and the automatic animation of the creatures behavior. The virtual creature is regarded as an autonomous agent that has sensors, actuators, and controllers. For controlling the virtual creature, an artificial neural network and a central pattern generator are adopted as the controllers. An optimization algorithm is introduced for training the controllers. Numerical experiments prove that the virtual creature realizes effective motions (walking behavior) to reach the destination and to avoid obstacles and other creatures.

Behavior emergence of virtual creature that can live in different environments

In this study, we aim at evolving autonomous virtual creatures which have complex shapes in a complex environment. We implement a basic physics law and fluid influences with a virtual environment and evolve artificial creatures in plural environments such as on the ground, and in the water. The model, which is evolved in a different environment, obtains effective different moving behaviors, walking and swimming in each environment.

A study on autonomous animated robots: Anibots

In this paper, we demonstrate a design of autonomous virtual creatures (called animated robots: Anibots in this paper) and develop a design tool for animated robots. An animated robot can behave autonomously by using its own sensors and controllers on three-dimensional physically modeled environment. The developed tool can enable us to execute the simulation of anibots on physical environment at any time during the modeling process. In order to simulate more realistic world, an approximate fluid environment model with low computational costs is presented. It is shown that a combinatorial use of neural network implementation for controllers and the genetic algorithm (GA) or the particle swarm optimization (PSO) is effective for emerging more realistic autonomous behaviours of animated robots.

A Study on Artificial Neural Network Structures for Agent Learning

Artificial neural networks which have been used for agent learning have mostly employed back-propagation and recurrent neural networks. We, however, have observed that there exists another network structure in life—a small world network, which is used by C-elegance, a kind of eelworms. We examined not only the performance of the small world network but that of a regular graph network and a random graph network. We applied these three networks to agent learning problems, and when we compared them with back-propagation and recurrent neural networks, it became clear that in the case of small world network structures, it has the same or even better performance as compared to back-propagation and recurrent neural networks despite a lower number of synapses.