الفهرس 
Chapter 1: IntroductionOnly 14 pages are availabe for public view
 |  | from | 130 |  |  |
| | | from | 130 | | |
Abstract
Multi-Agent Systems are used in a wide range of applications such as e-commerce, simulation, robotics, traffic control, manufacturing, health care, and Cloud Computing. The complexity of many tasks arising in these domains makes them difficult to solve with preprogrammed agent behaviors. Agents instead need to discover a solution on their own, using learning. The heart of the problem is how agents will learn the environment independently and then how they will cooperate to achieve the system goals. Furthermore, how the agents could coordinate and decide in order to achieve these goals. The objective of this study is to answer these commonly asked questions from the machine learning perspectives.
Multi-Agent Learning is not merely a matter of “straight” learning, but a matter involving complex patterns of social interaction and cognitive processes, which leads to complex collective functions. Many of the techniques developed in machine learning can be transferred to settings where there are multiple, interdependent, interacting learning agents, although they may require modification to account for the other agents in the environment. Furthermore, Multi-Agent Systems present a set of unique learning opportunities over and above single-learner machine learning. In a Multi-Agent System, an agent is always acting in the context of other agents, and so it must adapt its plans according to its expectations of the others.
Considering these interlinked issues of uncertainty and coordination, this thesis attempts to build upon existing algorithms for learning and decision making under uncertainty, including explicit models of other agents. A framework for interactive Multi-Agent Learning is proposed. Then, two modified machine learning algorithms are provided for collaborative agents learning, namely: Q-Learning and Multiply Sectioned Bayesian Network. The main case study applied in experiments is the Job Scheduling problem in Cloud Computing Systems. A model for multi-agent based Job Scheduling in Cloud Computing is proposed. Then, several jobs scheduling scenarios are experimented with the proposed framework and the two proposed machine learning algorithms showing the feasibility and improvements to the case-study performance.
First, an evaluation study is provided for Cloud Simulators. The scope of the study is to find the weak points of the existing Cloud Simulators for improvements. This thesis focused on the Job Scheduling concepts. Second, a Multi-Agent based model is proposed for the problem of Job Scheduling in Cloud Computing Systems. The proposed model improved the end-to-end utilization by 30% and the delay performance by 10%. The cause behind that is the shared blackboard pattern where agents store their information about the available resources. Third, a framework for Machine Learning in Interactive Multi-Agent Systems is proposed for a novel metaphor of interactions between agents. The proposed framework works on sharing information within the learning process among agents in the team. Proposed modifications to Q-Learning algorithm are provided following the proposed interactive framework and results are showing great improvement in the performance of the whole system. The mathematical basics are presented following the proposed behavior. Improvements provided include 2 times improvement in average reward received and 80% improvement in the number of trails to reach the goal. This is because agents were being able to get recommendations from other agents about their next actions that maximize their overall performance. At the end, another proposed and modified Multiply-Sectioned Bayesian Network algorithm is used for meta-scheduling problem in Job Scheduling in Cloud Computing. The proposed algorithm is saving computation power and storage for the meta-scheduling problems by 60%. Also, it improves the search time for a best fit machine by 40%. The proposed modifications to the algorithms works like a router that routes the request to the group of machines that has the highest belief to get the request assigned successfully taking in consideration the available resources and utilization.
The results of this work is believed to be one step towards enhancing existing machine learning algorithms for Multi-Agent Systems in dynamic environments. Moreover, there is a possibility of employing the proposed framework and algorithms to other applications areas of e-Health and data analysis with some modifications.