DISTRIBUTED AND APPROXIMATE REASONING

Products

Industries-of-the-Future

Papers:

Process Management

Distributed/Approximate
     Reasoning

Energy Systems

Working with limited domain data and ambiguous knowledge

Limitations of Data Mining

 Data mining is only the first step on the road to more effective business operations. Although mined business knowledge is valuable, it is limited for a number of reasons:

1. The extracted knowledge represents only what was in the database. Therefore, we cannot be sure that it can be used in all circumstances. Data analysts approach this problem by trying to make their database as representative of business as possible. They gather and analyze large quantities of data to improve the data's relevance to future business requirements.However, this is very expensive and having more data does not guarantee better business knowledge.
   
   
2. When working with imprecise and inconsistent databases, the knowledge discovered from data may represent only a small part of the business operations.Ideally, this problem should be addressed at the data mining stage by selecting a data mining tool specifically designed to work with imprecise and inconsistent data. However, if large inconsistencies or high imprecision characterizes your data, you will have to use additional tools in order to apply the extracted knowledge effectively.
   
   
3. The knowledge from databases represents only what has already occurred. Today’s business environments are rapidly changing and strategies used in the past may not be as applicable to the new business environment.

Intelligent Agents

As competition becomes fiercer the efficient development of business knowledge and strategies will become even more important. However, as discussed above, solving the problem of limited or obsolete business knowledge is time consuming and expensive.REDUCT & Lobbe Inc. answers this problem by supplying intelligent agents” to do the work for you. Intelligent agents can analyze your historical data and synthesize new feasible business strategies.  This allows you to adapt to new situations or business environments without the expense of updating and re-analyzing large databases, and without the problem of rapid obsolescence of your business strategies.

Intelligent agents are small programs that have enough reasoning ability to perform one well-defined task. A number of agents are already being used by business for information retrieval and management.One example is the Topic (Verity) series of software tools that search for, retrieve and filter information on the Internet. Hewlett-Packard’s paperless wage review system uses agents to automatically start the employee review process and send out forms (by e-mail or fax) to the appropriate managers.Personal Information Agent and Advanced Information Agent (J&T Associates) sort mail, copy relevant information into a user’s personal database, and automatically reply to e-mail.All of these applications use agents to manipulate existing information.

Distributed Reasoning

 Distributed reasoning provides the basis for dealing with uncertainty in knowledge assertion and for building hypotheses that describe our knowledge imprecisely. In distributed reasoning, the agents perceive the requirements as vague statements, which, in the language of agents, are expressed as approximate specifications. The term “distributed” means that each agent can work on different aspects of the problem. These agents are like experts who have the knowledge required to perform tasks such as problem analysis, decomposition, synthesis and solution evaluation, using their respective expertise. They can also form elementary teams of agents with well-defined goals. 

REDUCT’s distributed reasoning technology uses a system of cooperative, intelligent agents to form a powerful set of problem-solving tools which can actually create new strategy designs from existing knowledge (business rules). “Cooperative” means that they can exchange information and meet local as well as global requirements. The distributed structure allows the agents to break the strategy design problems into smaller tasks that can be more effectively solved. Since no agent can solve the problem by itself, the agents must cooperate and build common solutions.

 Each agent decides what knowledge components it needs in order to work on the design problem given to it, and if it does not have the components itself the agent requests them from subagents. This request is in the form of an approximate specification that defines the requirements as well as the evaluation criteria. The subagent uses the evaluation criteria to determine how well a prospective knowledge component meets the requirements. Components, which pass the evaluation, are submitted to the requesting agent as recommendations. The requesting agent then synthesizes the recommended component knowledge into one or more feasible strategy designs.

Application Advantages

Distributed reasoning technology lends itself to a number of business and industrial applications in which the quality and number of available strategies could be expanded.  Some examples are:

• target marketing, credit analysis, real estate appraisal, human resources, fraud analysis, etc.
  
• optimizing loop and supervisory control, developing new designs, scheduling and dynamic task assignment.

In distributed reasoning, each intelligent agent works on one clearly defined problem, so increasing complexity does not adversely affect the quality of the strategy designs. The knowledge derived by distributed reasoning increases the scope and usefulness of knowledge discerned by data mining and the strategies derived will not become quickly obsolete as the environment changes.

Limitations of Knowledge

The use and application of decision support systems has to reflect the inexactness and ambiguity of the historical data and knowledge rules derived from them. Specifically it has to account for:

1. What is measured is often ill defined.  Data are linked to the representation (simplification, aggregation, etc.) of the system.
   
   
2. The business/process rules derived from the historical examples will provide incomplete knowledge about data dependencies and relationships.
   
   
3. The rules are only an approximate representation of some aspects of business.  This means that in addition to omissions, simplification and aggregation, there are factors which have not been captured by the rules.
   
   
4. Empirical strategies allow us to make decisions only a fraction of the time because they are incomplete, i.e. describe only past experiences.  This means that we will have to use approximate reasoning methods to make inferences.
   
   
5. The future almost always conceals something unpredictable or indeterminable.    This may be caused by environmental (business) changes, but more often it is because complex systems cannot be determined in a comprehensive way.  Thus, in addition to the external factors (environment) responsible for the uncertainty, there are internal factors which become prominent under different conditions.
   
   
6. We will also need some criteria to decide when successful past strategies are no longer applicable, i.e. new strategies are needed. 

Approximate Reasoning

Approximate reasoning is an entirely new technology for intelligent decision support that overcomes many of the difficulties in modeling data and decision making based on past experiences.The technology helps to obtain answers to questions involved in a decision-making process. These answers clarify the decision and increase coherence between the decision process and goals. The approximate reasoning methodology combines a number of analytical approaches including Artificial Intelligence methods. The key ambiguities addressed by approximate reasoning methods are the limitations of the mono-criteria approach in classical decision theory such as value, utility, efficiency or entropy functions. These limitations are overcome by the use of multi-criteria methods based on “meta” concepts describing associations, behavior, and preferences.

 Approximate reasoning recognizes that all decisions encounter many uncertainties arising from the multiplicity of variables and factors, which are impossible to control completely. The discrepancies, the purpose of the decisions, certain aspects of the consequences, etc., all can be the cause of uncertainties. Therefore, the objective of approximate reasoning is not to know the “best possible” decisions, but to develop the “most suitable” decisions.

As implemented in REDUCT Decision Suite (RDS), the technical benefits and advantages offered by approximate reasoning arise from the high level reasoning capabilities of approximate reasoning and from the capability to work with ambiguous data and knowledge. Approximate reasoning capability parallels, to a great extent, the human capability to reason and perform multi-criteria decisions. As a result, RDS can perform adaptive optimization strategies for very complex, not well-understood business processes. RDS can adapt to changing requirements for decision making by updating its decision strategies.

Application Advantages

Approximate reasoning provides a unique methodology for building decision support systems, namely it:

1. provides a proactive approach to decision making by directing the decision maker towards the use of different normative strategies;
   

2. reduces the effort required to explore strategy change.

It provides all the benefits of traditional modeling tools and decision support systems, yet it also offers new and powerful capabilities.

REDUCT & Lobbe Technologies Inc.
P.O. Box 800,  186 - 8120 No.2 Road., Richmond, BC,  Canada  V7C 5J8
ph: (604) 275-3711   fax: (604) 275-3715  email: dispatch@reduct.com