ECliPS Description

ECliPS is a knowledge management system designed to organize and process data and information about the climate system. Due to existence of so-called "teleconnections," or linkages between climate anomalies over great distances, an accurate seasonal forecast in one area is practically impossible without taking into account climatic processes in other areas of the world, as well as external factors, such as solar activity. This leads to a huge information overload, which is one of the main problems in climate forecasting. In this situation, the organization of information, along with reconciliation and connection of evidence and ideas, plays a very important role. 

There are two different ways to deal with the information overload. One large group of methods tries to resolve this problem by reducing the dimensionality of the system. This group includes principal component analysis, singular value decomposition, multidimensional scaling and other methods. These methods are very popular in climatology. Indeed, they can be useful in a number of situations, although there is often a problem of interpreting the results. Another important drawback of those methods is that they do not preserve information about the relationships between climate variables.

An alternative approach to the problem of information overload is to use those methods or tools that can help manage information in such a way that only the information relevant to the problem or question at hand is provided to the user at any given point of the analysis. This approach can be realized in the form of an expert system, decision support system or information management system. Climate Logic advocates such a system approach to climate forecasting, when the entire climate system is taken into consideration, even at the expense of some simplification.

The first climatic expert system called CESNA was developed by Rodionov and Martin (1996). For several years it was used in forecasting of winter atmospheric circulation and temperature in the North Atlantic and Europe (Rodionov and Martin, 1999). These experimental forecasts demonstrated the advantages of symbolic processing, which allowed taking into account both quantitative and qualitative information, and thereby catching many aspects of climate variability that were escaped from other conventional methods. Unfortunately, it was written in a rare language (KnowledgeMan/GURU) for DOS operating system and was somewhat cumbersome to use. 

Unlike CESNA, ECliPS is written in VB.NET with the use of several off-the-shelf Microsoft products: Word, Excel, Access, and Visio. It is more user-friendly and much easier to use than CESNA. It streamlines the inference process and makes it more transparent to the user.

Major components

The major components of ECliPS are: Data Explorer, Rule (or Knowledge) Explorer, Inference Engine, Graphical Interface, Search and Reporting Facilities. In many respects ECliPS is similar to an typical expert or decision support system, but unlike other commercial expert systems, both the knowledge presentation and inference process in ECliPS are more transparent to the user, and it is much easier to use. In this section we describe the first two components, Data and Rule Explorers, which help to manage the data and knowledge bases, respectively. Other components will be briefly discussed in the next section that explains how the system arrives to its conclusions.

Data explorer

Fig. 1. Data explorer

The Data Explorer (Fig. 1) organizes information about climate variables based on geographical hierarchy. The user can easily create his/her own geographical domain with the necessary level of details. The data for each climate variable is kept in a separate Excel file and the descriptive information in a Word file.

Fig. 2. Edit form for a climate variable.

Each variable has four attributes: name, region, season/month, and base period against which the anomalies are calculated (Fig. 2). The data can be presented as a set of numerical values and/or classes (categories), including fuzzy sets, with the corresponding degrees of membership (Fig. 3). The Data Explorer also allows the user to see a list of rules, in which the selected variable is used either as predictand or predictor (Fig. 4). This is a quick and convenient way to estimate the role of the variable in the knowledge domain, which in our case is (ultimately) the entire climate system. It also significantly simplifies the construction of a complex forecasting project with multiple links.

Fig. 3. Data template.

Fig. 4. A list of rules that describe factors affecting the timing and magnitude of ENSO.

 

Rule explorer

Fig. 5. Rule Template.

The domain knowledge in ECliPS is presented as a set of rules. Most commonly, the rules are in the IF-THEN form. For example, a rule may look like

IF ENSO event = warm,
AND Aleutian low circulation type = W1,
THEN winter SAT in the Bering Sea = above normal;
CF = 50.

Here CF is the confidence factor for the rule. Despite the simplicity, this type of rules is very powerful and versatile in expressing  relationships between climate variables.

Fig. 6. Rule edit form.

It is important to note that the data and code for each rule are placed in a separate Excel file. This provides substantial  flexibility in preparing a rule for the knowledge base. In most cases it would be suffice to use a provided template (such as the one in Fig. 5), where the user can simply fill in the table. The advanced users, however, who are familiar with Visual Basic for Excel, can write their own code. Another advantage of this rule information storage is that the user can easily experiment with each rule separately and develop a better feeling of confidence in it.

Fig. 7. Rule Explorer.

Common management tasks, such as inserting new rule, editing (Fig. 6), deleting, sorting, etc., are performed using the Rule Explorer (Fig. 7). The number of variables in the IF part of a rule is unlimited.

 

How it works

Fig. 8. The influence diagram showing the predictors for ENSO.

With the help of Data and Rule Explorers the user can insert variables and rules into a project and construct an influence diagram. As an example, Fig. 8 shows the influence diagram for El Niño -Southern Oscillation (ENSO). 

When the influence diagram is prepared, the user may run the project in the forecast or hindcast mode to infer the value of the target variable in a given year (Fig. 9). During this process, the system asks for the information about the variables in the terminal nodes of the diagram (Fig. 10).

Fig. 9. Running the project in the forecast mode.

To facilitate the answer to those questions, the user is provided with the access to the data and descriptive information about the variable and related rule. The user can also search for any other pertinent information (Fig. 11).

Fig. 10. Questions about the variables in the terminal nodes.

The evidence from different sources is combined using the CF algebra. In addition, the Bayesian inference technique may be added later. When all the evidence is collected, a forecast for the target variable is issued, showing its categories (classes)  with the associated CFs of their occurrences. 

The user can also open the Custom Property window (Fig. 12) and check the information about individual variables and rules, or open the report that traces the logic behind the forecast (Fig. 13).

Fig. 11. Search form.

Fig. 12. Custom Properties window.

Fig. 13. Report showing the inference process.