Merging Neural Networks in a Multi-Agent System

Though various interesting research problems have been studied in the context of learning agents, few researchers have addressed the problems of one, knowledgeable, agent teaching another agent. Agents can do more than share training data, problem traces, learned policies. In particular, we investigate how an agent can use its learned knowledge to train another agent with a possibly different internal knowledge representation. We have developed an algorithm that can be used by a concept learning agent, the trainer, to iteratively select training examples for another agent, the trainee, without any assumptions about its internal concept representation or learning algorithm. We present results on artificial and real-life concept learning problems using our Agent Teaching Agent (ATA) framework where instance based and decision tree learners are used as trainer and trainee agents.

This paper describes techniques for integrating neural networks and symbolic components into powerful hybrid systems. Neural networks have unique processing characteristics that enable tasks to be performed that would be difficult or intractable for a symbolic rule-based system. However, a stand-alone neural network requires an interpretation either by a human or a rulebased system. This motivates the integration of neural/symbolic techniques within a hybrid system. A number of integration possibilities exist: some systems consist of neural network components performing symbolic tasks while other systems are composed of several neural networks and symbolic components, each component acting as a self-contained module communicating with the others. Other hybrid systems are able to transform subsymbolic representations into symbolic ones and vice-versa. This paper provides an overview and evaluation of the state of the art of several hybrid neural systems for rule-based processi...

2009

Because of the big complexity of the world, the ability to deal with uncertain and to infer "almost" true rules is an obligation for intelligent systems. Therefore, the research of solution to emulate Inductive Reasoning is one of the fundamental problem of Artificial Intelligence. Several approaches have been studied: the techniques inherited from the Statistics one side, or techniques based on Logic on the other side. Both of these families show complementary advantages and weakness. For example, statistics techniques, like decision trees or artificial neural networks, are robust against noisy data, and they are able to deal with a large quantity of information. However, they are generally unable to generate complexes rules. On the other side, Logic based techniques, like ILP, are able to express very complex rules, but they cannot deal with large amount of information. This report presents the study and the development of an hybrid induction technique mixing the essence of statistical and logical learning techniques i.e. an Induction technique based on the First Order Logic semantic that generate hypotheses thanks to Artificial Neural Networks learning techniques. The expression power of the hypotheses is the one of the predicate logic, and the learning process is insensitive to noisy data thanks to the artificial neural network based learning process. During the project presented by this report, four new techniques have been studied and implemented: The first learns propositional relationship with an artificial neural network i.e. induction on propositional logic programs. The three other learn first order predicate relationships with artificial neural networks i.e. induction on predicate logic programs. The last of these techniques is the more complete one, and it is based on the knowledge acquired during the development of all the other techniques. The main advance of this technique is the definition of a convention to allow the interaction of predicate logic programs and artificial neural networks, and the construction of Artificial Neural Networks able to learn rule with the predicate logic power of expression.

Communication between cooperating agents has been shown in research to enhance the learning process of agents and their ability to complete tasks successfully in a group environment. However, an environment where we have many independent interacting components (agents), with each affecting the decision processes of the others, is essentially a complex system. Studying the behavior of the inter-agent communication in a complex system is difficult, particularly as the interaction between the agents may also affect the communication type and level. However, modeling the agent communication as a Neural Net may provide a convenient way to view such communication and further research it's effect on the learning process.

In a vision system, every task needs that the operators to apply should be {\guillemotleft} well chosen {\guillemotright} and their parameters should be also {\guillemotleft} well adjusted {\guillemotright}. The diversity of operators and the multitude of their parameters constitute a big challenge for users. As it is very difficult to make the {\guillemotleft} right {\guillemotright} choice, lack of a specific rule, many disadvantages appear and affect the computation time and especially the quality of results. In this paper we present a multi-agent architecture to learn the best operators to apply and their best parameters for a class of images. Our architecture consists of three types of agents: User Agent, Operator Agent and Parameter Agent. The User Agent determines the phases of treatment, a library of operators and the possible values of their parameters. The Operator Agent constructs all possible combinations of operators and the Parameter Agent, the core of the architecture...

Journal of Intelligent Manufacturing, 1992

This paper introduces a recurrent network for mapping of interclass members without performing a learning process. This allows a member of class A to be mapped to a member of class B. Given sample members of each class, a backpropagation network is trained to form the corresponding class boundaries. Upon completion of the training process, the weights obtained are used in a recurrent network which performs the interclass member mapping without any further training. This mapping is achieved as the recurrent network evolves in time. The initial state of the network is mapped to its equilibrium state. The interclass member mapping network (IMMN) has many applications in self-correcting systems. In this paper, the IMMN is developed to represent two classes, namely class B (for instance, a class for representing members with desirable and correct features) and class A (members with incorrect features). The self-correcting aspect of the network is demonstrated by mapping a member from class A to a member of class B. Two examples are given. One represents a case of linearly separable classes. The other explores an application example, where two categories are used, namely, poorly and well-designed manufacturing parts. Given a poorly designed part, the network will suggest corrections resulting in a well-designed part. This example has nonlinear decision regions and shows the generalization capability of the network.

A major challenge in multi-agent reinforcement learning remains dealing with the large state spaces typically associated with realistic multi-agent systems. As the state space grows, agent policies become more and more complex and learning slows down. Current more advanced single-agent techniques are already very capable of learning optimal policies in large unknown environments. When multiple agents are present however, we are challenged by an increase of the state-action space, exponential in the number of agents, even though these agents do not always interfere with each other and thus their presence should not always be included in the state information of the other agent. We introduce a framework capable of dealing with this issue. We also present an implementation of our framework, called 2observe which we apply to some gridworld problems. Furthermore we demonstrate that our approach is capable of transferring its knowledge to new agents entering the environment.

2008 IEEE International Joint Conference on Neural Networks (IEEE World Congress on Computational Intelligence), 2008

Since machine learning has become a tool to make more efficient design of sophisticated systems, we present in this paper a novel methodology to create powerful neural network controllers for complex systems while minimising the design effort.

TIntelligent multi agent systems have great potentials to use in different purposes and research areas. One of the important issues to apply intelligent multi agent systems in real world and virtual environment is to develop a framework that support machine learning model to reflect the whole complexity of the real world. In this paper, we proposed a framework of intelligent agent based neural network classification model to solve the problem of gap between two applicable flows of intelligent multi agent technology and learning model from real environment. We consider the new Supervised Multilayers Feed Forward Neural Network (SMFFNN) model as an intelligent classification for learning model in the framework. The framework earns the information from the respective environment and its behavior can be recognized by the weights. Therefore, the SMFFNN model that lies in the framework will give more benefits in finding the suitable information and the real weights from the environment wh...