Juan R. Velasco

Analysis and Design of Multiagent Systems Using MAS-Common KADS

This article proposes an agent-oriented methodology called MAS-CommonKADS and develops a case study. This methodology extends the knowledge engineering methodology CommonKADS with techniques from objectoriented and protocol engineering methodologies. The methodology consists of the development of seven models: Agent Model, that describes the characteristics of each agent; Task Model, that describes the tasks that the agents carry out; Expertise Model, that describes the knowledge needed by the agents to achieve their goals; Organisation Model, that describes the structural relationships between agents (software agents and/or human agents); Coordination Model, that describes the dynamic relationships between software agents; Communication Model, that describes the dynamic relationships between human agents and their respective personal assistant software agents; and Design Model, that refines the previous models and determines the most suitable agent architecture for each agent, and the requirements of the agent network.

Multiagent-based control systems: A hybrid approach to distributed process control☆

Abstract This paper presents a general architecture and a platform developed to implement distributed applications as a set of cooperating intelligent agents. It also shows how this architecture has been used to implement a distributed control system for a complex process: the economic control of a fossil-fuel fired power plant. Agents in this application encapsulate different distributed hardware/ software entities: neural and fuzzy controllers, a data-acquisition system, presentation manager, etc. These agents are defined in ADL (Agent Description Language), a high-level specification language, and interchange data/knowledge through service requests using a common knowledge-representation language.

Genetic-based on-line learning for fuzzy process control

This paper deals with the problem of continuous learning in process control. Conventional machine learning applied to process control tries to obtain control rules from an historic data file or a model. However, these learned rules may be useless if the real process changes, and this is not unusual. To try to solve this problem, genetic algorithms can be used in a continuous learning environment. However, genetically generated rules do not guarantee that they are good enough to control the process. New rules should be tested before their insertion into the knowledge base: this is the function of Limbo. Limbo is a special place where rules can be tested in real situations before being used. This paper shows how Limbo can be used to improve continuous learning.

MIX: a general purpose multiagent architecture

The MIX multiagent architecture has been conceived as a general purpose distributed framework for the cooperation of multiple heterogeneous agents. This architecture, starting from previous work in our group on multiagent systems, has been redesigned and implemented within a research project investigating a particular class of hybrid systems: those integrated by connectionist and symbolic components. This paper describes in some detail the principal concepts of the architecture: the network model and the agent model. Around these models, a set of languages and tools have been developed. In particular, an Agent Description Language (MIX-ADL) has been designed to specify agents declaratively in a hierarchy of classes.

Inductive Learning Applied to Fossil Power Plants Control Optimization

Abstract This paper introduces a couple of non classical control systems with learning capabilities, applied to optimize the heat rate on a fossil power plant. Both systems suggest actions to be performed by a plant operator, with an unique goal: to get heat rate as low as possible. Different inference and learning methods have been developed: a classical expert system with a genetic algorithms module to create new rules and a fuzzy control based system where new rules are created using similar methods. Both systems have been installed and tested in a real power plant, assisting operator to decide about the best actions.

A Proposal for Meta-Learning Through a Multi-Agent System

The meta-learning problem has become an important issue in the recent years. This has been caused by the growing role of datamining applications in the global information systems of big companies which want to obtain benefits from the analysis of its data. It is necessary to obtain faithfull application rules that guide the datamining process in order to achieve the best possible models that explain the databases. We follow an inductive approach to discover these kind of rules. This paper explains the MAS-based information system we use for mining and meta-learning, and how the scalability problem is solved in order to support a community of many software agents.

A Control Architecture for Optimal Operation with Inductive Learning

Abstract This paper describes an architecture for expert process control for those processes where learning is needed. Machine learning let the system to discover new rules to act over new circumstances and improve process performance in known areas. The architecture is open enough to allow developing such different control systems as fuzzy or classical logic based ones. The real application described in section III shows the capability of two different process control systems designed under this scheme. The rule evaluation mechanism let the system to modify the credibility of every rule which is able to be used. This algorithm makes possible the existence of a limbo: a place where induced new rules are tested, before to be included into the knowledge base.

Genetic Fuzzy C-Means Algorithm for Automatic Generation of Fuzzy Partitions

Automatic knowledge generation for Fuzzy Rule Based Systems comprises two main tasks: rules generation and the definition of the semantics of the linguistic variables applied by the rules. The paper focuses on the problem of semantics’ definition by using the Fuzzy c-means algorithm. Fuzzy c-means is a clustering algorithm widely used but with several drawbacks, such as the dependence of the results on the initialization and the need to predefine the number of clusters to be generated. To solve the first problem different authors have used genetic algorithms, defining the genetic fuzzy c-means (GFCM) clustering algorithms. In addition, the paper will present a GFCM clustering algorithm which also finds the suitable number of clusters.

Computational intelligence in process control

In the past years, the Intelligent Control Group at the Universidad Politecnica de Madrid has been developing architectures and applications for complex process control based on artificial intelligence technologies. We present some results obtained and comments about the work under development. The most active issues in these days are the specification of an intelligent control system design methodology based on a task-technology map and the integration of specific technologies to obtain synergistic effects. We study the application and integration of the following technologies: fuzzy control and validation, neurocontrol, neuromodelling and genetic algorithms for optimization and control adaptation.<<ETX>>

Multiagent-Based Control Systems: An Hybrid Approach to Distributed Process Control†

Abstract In this paper a general architecture and a platform developed to implement distributed applications, as a set of cooperating intelligent agents, is presented. Second, it will be shown how this architecture has been used to implement a distributed control system for a complex process: the economic control of a fossil power plant. Agents in this application encapsulate different distributed hardware/software entities: neural and fuzzy controllers, data acquisition system, presentation manager, etc. These agents are defined in ADL, a high level specification language, and interchange data/knowledge through service requests using a common knowledge representation language.