Descripción: |
A Multi-Agent system, a loosely coupled network of solvers which interact
to find a solution to a problem beyond individual capabilities and knowledge
[29], is a common notion in the literature with application to a broad range
of problems. One approach in the design and application of such systems is
machine learning. Machine Learning deals with the design of programs which
take advantage of data, examples and experience to improve accuracy or
performance [16]. Specifically, the area of Machine Learning algorithms that
deals with Multi-Agent systems is known as Ensemble Methods. Common problems
addressed by these techniques are Classification and Prediction. When designing
a Multi-Agent system using Ensemble Methods, 4 different stages can be
identified: 1. Pre-processing 2. Partition 3. Training 4. Post-processing
These 4 stages are independent and have different goals. Preprocessing refers
to preparing the data in order to improve learning efficacy. Partition refers
to dividing the data among the different agents. Training corresponds to
the process where the data is used to learn how to solve the problem at hand.
Finally post-process includes techniques to analyze or modify the learning
results. At the end of this process, enough information has been learned
to attempt to solve unseen problems of the same type. For each stage, there
exist multiple techniques that may be applied which might be proper for one
sort of problem but not for the other. One specific combination might be
more adequate than other and finding an optimal combination is an aspect
of our research.
To habilitate experimentation, we have designed a framework in which different
interchangeable components may be connected and different Multi-Agent Systems
created. These systems may then be exported through XML to be used in other
applications. Using this framework, a comparative analysis on the different
stages was performed, and an ensemble based solution was applied for the
HLA multi-classification problem [20], for which our research represents
the first attempt of applying machine learning techniques. |