Bayesian network system using neural networks for representation of
conditional probabilities is introduced. In order to apply Bayesian network to
real world applications, for example office-conversant robot, continuous real
value or discrete multi-value should be treated easily in the system.
Then we propose neural-bayesian network modeling, which uses multi-layer
parceptrons to represent conditional probability in each node.
In this paper, learning scheme is also shown through mobile robot navigation
task.
Bayesian networks are studied as graph representation for probabilistic
reasoning of uncertain event. In real world problems,
multi-dimensional discrete values and continuous values should be handled
in many cases. In order to compute probabilities of such information naturally,
a Bayesian network model using multi-layer perceptrons is proposed.
This model is consider as a large neural network constructed of many
two layer perceptrons on directed acyclic graph structure.
A simulator program written in JAVA language is also developed to evaluate
this model through probabilistic reasoning for real world problems.
Real world computing, which is robust, reliable, and adaptive computation
mechanism in the real word environment, is studied in our projects.
In this paper, an intelligent autonomous mobile robot system is introduced
as an example of real world computing.
The system learns the map of real office environment through simple dialogue
with a human teacher, and navigates to the correct place in the environment
under many kind of uncertainty.
In order to treat uncertain information, incorrect sensor data with
noise, unexpected change of environment and so on, integration technique
using Bayesian network is applied.
For this system, it is important
to make an anytime estimation of the probability values, so a Monte Carlo
method is applied.
Finally, this framework is applied to the self position estimation in the navigation task.
In RoboCup competition, to realize flexible adaptative ability is an important issue as with a standard problem in AI. For this purpose, identification of an opponent's strategy is studied in this paper. Some statistical methods, Bayesian estimation method, information maximizing strategy and the multi-armed bandit theory, are introduced according to the RoboCup regulation. Finally, an observer agent is designed to implement these features. This program is written in JAVA with multi-threading to realize both reactive response and deliberative planning.
In this paper, "Focal Point"(Sheiling 1960) is discussed as a method of sharing information without explicit communication between cooperative multiple autonomous robots. We consider cooperation behavior as multiple robots arriving at the same target object at the same time. The question is how to dispatch multiple robots to a common object reactively. We provide a framework that embodies the focal point concept in the context of cooperative multiple robots, and some focal point algorithms are applied. In order to evaluate cooperative performance quantitatively, multiple autonomous robots simulator(MARS) has been developed by our group. In two dimensional environment, each robot move around and observes objects using an eye-sensor with limited field of view. So, a robot can only observe part of all objects until it explore completely. For this reason, some focal point algorithms using complete observation are not sufficient in this problem. A simple focal point algorithm, which doesn't need complete observations is proposed. Using our simulator, some experimental results show that this algorithm can improve cooperation performance efficiently. Then, some other experiments show that sharing observing experience with other robots can also solve the requirement about complete observations.