WO2004104919A2

WO2004104919A2 - Expert system for evaluating an information gain with regard to a state of a system

Info

Publication number: WO2004104919A2
Application number: PCT/EP2004/002225
Authority: WO
Inventors: Joachim Horn; Marco Pellegrino
Original assignee: Siemens Aktiengesellschaft
Priority date: 2003-05-22
Filing date: 2004-03-04
Publication date: 2004-12-02
Also published as: WO2004104919A3

Abstract

The invention relates to an expert system for evaluating an information gain with regard to the state of a system, comprising: a means for determining an item of state information of the system and; an evaluation unit for evaluating the item of state information. According to the invention, an item of information (ζ1) concerning an actual state of a system is determined and at least one item of additional information ((ζ2) concerning the actual state of the system is determined. A first forecasted state (Pr1((ζ1)) of the system is determined from the item of information, and a first probability (P1(St)) is determined with which the first forecasted state matches the actual state of the system. A second forecasted state of the system (Pr2(akl U (ζ)) is determined. A second probability (P2(St)) is determined with which the second forecasted state matches the actual state of the system. A difference (Dp1p2) between the first and the second probabilities is determined, and this difference is used as an evaluation measure for an information gain with regard to the system based on the item of additional information, and the determining means is controlled according to said evaluation measure.

Description

description

Expert system for evaluating a gain in information about the state of a system.

The invention relates to an expert system for evaluating an information gain about a state of a system.

With automatic recognition systems, information about a system to be recognized must be recorded. The information can be used to provide a forecast or an assessment of the state of the system. The will practice; the system available information is usually presented in the form of statistical state variables. Examples of such variables are physical quantities in a detection system for detecting diseases, e.g. Color properties of a characteristic appearing on a body, body temperature, blood pressure or other quantities that can be measured on a body.

Known recognition systems use methods in which models of the system to be recognized are generated which enable a meaningful and quick evaluation of information. A number of methods for the evaluation of medical are in particular for medical recognition systems

Known images and extraction of disease information from the images. The symptoms of a patient are usually determined using image recordings. The information is usually present in an image in such a way that it is read in the course of a scanning process, e.g. with a camera, and finally can be evaluated in one continuous run using a method suitable for recognition.

A method is known from [1] in which images are segmented into specific areas, in the segmented areas relevant information of non-relevant (superfluous) gen) information is separated and the relevant information is then extracted and evaluated from the segmented areas. Gray values, contour lines or texture information are read from the images. A background pattern, for example, is understood here as superfluous information, which should be separated from a relevant foreground feature of the image. By eliminating the background pattern, less data can be processed so that the image can be classified more quickly.

For the evaluation of the features contained in the images, certain statistical methods, such as, for example, from [2], are known which use neural networks or clustering models for recognizing and classifying the objects contained in the images.

For automatic detection processes in which a state of a system cannot be detected in one go, e.g. When determining a patient's initially unknown illness, methods are used which carry out the gradual accumulation of information until a reliable prognosis of the actual state of the system is possible. These systems must be highly adaptive to enable quick diagnoses from very different physical or physical symptoms, such as different fever temperatures that are individual for a specific disease, i.e. can vary from patient to patient.

The relevance of information about a state of a system has so far been determined using variables such as entropy, variance, weight of evidence, utility or variables derived therefrom [3].

For example, a method is known from [3] in which the condition information of a system is based on an entropy measure Is evaluated. The emphasis is on minimizing the entropy value through the collection of information by means of suitable questions to a system to be recognized, or on reducing the uncertainty of the knowledge when collecting the information about the system. For an observation or information, an overall uncertainty regarding a statement or information about a state of the system derived from the observation is statistically evaluated.

For reasons of optimizing such an algorithm, consideration of costs / benefits should be taken into account when observing or procuring status information. The algorithm is also designed in such a way that the value of the total uncertainty or that of the entropy decreases after a question is asked. However, it has some disadvantages. Obviously relevant information that leads to an increase in the entropy measure, e.g. the question of vaccination against a disease whose symptoms are present is not taken into account. In this case, after a few steps during the execution of the algorithm, the accumulated information, consisting for example of symptoms A, B, and C of the system, which increases the likelihood of a certain disease, could no longer be asked the question about a vaccination, since this question would increase entropy.

Information about the system can at the same time increase the likelihood of a correct prognosis of one disease and decrease that of another, so that the entropy ate remains constant. Nevertheless, this information or the corresponding question has relevance for the final determination of the state of the system.

The relevance of a question is therefore not adequately described by the expected change in the entropy measure. Another disadvantage of the entropy-based method is that it is sensitive to numerical errors.

The invention is therefore based on the object of specifying a device in which information about a state of a system can be assessed more reliably and better.

The task is solved by an expert system and a method according to the respective independent claims.

The invention thus consists of an expert system for determining a status of a system, which comprises the following features: a means for determining status information of the system,

a control unit for controlling the agent,

- An evaluation unit for evaluating the status information, wherein - using the means

- Information about an actual state of the system is determined and

- at least one additional piece of information about the actual state of the system is determined, - using the evaluation unit

a first predicted state of the system is determined from the information,

a first probability is determined with which the first predicted state matches the actual state of the system,

a second predicted state of the system is determined using the combination of the information and the additional information, a second probability is ascertained with which the second predicted state corresponds to the actual state of the system, - A difference between the first and the second probabilities is determined and this is used as an evaluation measure for an information gain about the system through the additional information, and - using the control unit, the mean is controlled depending on the evaluation measure.

The processes carried out by the means and by the units are preferably carried out iteratively, so that a final determination of a state of the system can be achieved.

The expert system has the advantage that even information that would increase the entropy of knowledge in the state-of-the-art entropy-based method and would therefore not be used there is reliably evaluated by using the direct relationship between the first and the second probability can be. This enables rapid convergence to a determination of the actual state of the system.

A “system” can be an object, a grouping of interacting objects or a human body. The objects can be technical devices, human organs or structures within an electronic network, for example the Internet.

“Status information” means information about the status of the system.

“Additional information” means additional information about the state of the system.

“Actual state” is understood to mean the state of the system which is actually present and which is ultimately to be recognized or determined by the accumulation of state information. The means for determining status information of the system are preferably sensors. Alternatively, the means can also be implemented as an input or receiving unit for information, for example as a camera or a keyboard. The means are able to determine information about a state of the system and additional information about the state of the system.

The control unit is preferably a switching unit, the function of which is to activate the means for determining status information of the system.

The evaluation unit is preferably a computer program product, the function of which is preferably an evaluation of status information of the system.

It is preferred that the functions of the evaluation and control units are taken over by a computer.

It is also preferred that the evaluation and control units are controlled by a single computer program product and are each part of a computer.

The expert system preferably has a storage medium or a plurality of storage media in order to store state information of the system, modeled states of one or more systems and program code for controlling the evaluation and control units. The modeled states and stored state information of the system belong to a knowledge base of the expert system. Portable storage media such as CD-ROM or floppy disks can of course be used.

The expert system is preferably used as a technical aid for determining the condition of a human body. The determination is preferred a diagnosis, which is aimed for a body of unknown state. The means for determining status information of the system thus request, for example, a patient, for example via a screen, to provide information about their physical condition. This information is preferably recorded by the expert system using a keyboard.

The means and units mentioned in the expert system with their respective functions mentioned therein are used in the method for evaluating information gain about the state of a system.

It is advantageous if, in the expert system or in the method, the relevance of an expected state information, for example after a request for the state information, is determined. The relevance is assessed by the fact that the information gained by fulfilling the information requirement is a criterion of relevance. The relevance of the response to be expected can be a priority of the request for the status information.

An example of the above-mentioned request for status information is a question to a patient or the query of a sensor which can determine a status of a system.

The determination of status information about the system can therefore be implemented by requesting status information about the system.

In order to provide a more precise statement about the gain or the relevance of information, a cost / benefit analysis is carried out for the determination of the status information, an optimal request for information only being carried out when the benefits of the request, that is to say the possible the information gained through the request is greater than the cost of the request.

In order to enable a faster convergence to a final statement about the actual state of the system, a list of certain requirements for information in a memory is presorted from the current available information status to a smaller selection of the requirements. In the subsequent evaluation, the expert system has a smaller list of possible requirements available, so that a reduced computing effort is guaranteed.

The invention will be explained in more detail with reference to the drawing and exemplary embodiments. It shows

1 shows the course of a relevance assessment of status information via a system.

It is preferable to evaluate the information gained about the state of the system

a first probability function for the state of the system is determined from first information about the system, a second probability function is calculated from this first information and additional information,

a distance measure between the probability functions is calculated by summing a non-negative distance measure of the probabilities over all possible states of the system,

- The expected value of the distance measure between the probability functions is determined and this is used as an evaluation measure for gaining information about the system through the additional information.

FIG. 1 shows a sequence of evaluating information to be requested about the system, in which a stored te list 100 of possible questions q _k , starting with questions ql and finally with question qn, for the question with the highest relevance for the given level of information about the system. In this example, the information to be requested is understood as a question.

An a priori stochastic distribution (a model), or given information, or else an observation or observations φ 1, step 101, of the system to be diagnosed is assumed. The observation is carried out by the means for determining status information of the system. On the left in the figure is a column 100 in which a list of questions q _{k is} given. With the information φ 1, a prognosis Prl (φl), step 102, that the system is in a certain state is determined. First, the probability that the prognosis Prl (φ) corresponds to the actual state Pl (St), step 103, is determined by a computer or by the evaluation unit.

Step 104 in the figure indicates the question q1, which is for example asked of a patient. The question is first exposed in step 106 to an answer a _k ι from the system S (bottom left), which could be answered with “yes” or “no”, for example, or not at all. The answer a _i is an example of additional information φ2, which can take up a binary form 1 or 0 in a computer. After receiving the additional information φ2, a next forecast Pr2 (a _ki U φ) is determined in step 107, which uses the knowledge from a combination of the information φl and a _k ι or φ2. First, a second probability P2 (St) that the prognosis Pr2 (a _k ι) matches the actual state is determined in step 108. The difference between the first and second probabilities P1 (St) and P2 (St) is thus specified in step 105 D _p ι _p2 . It is a measure of the relevance of a question q. The value of this difference then becomes a database or forwarded to list 100, where either the question is actually asked to the patient if the difference is sufficient, or list 100 is searched for a further, more relevant question. All possible questions are exposed to an expected answer. To answer the question q, the control unit controls the means for determining the status information of the system as a function of the difference between the probabilities. The regulation can be used to obtain information to be received or an answer to a question

a _{ki is} the expected answer to a question to be asked q _k , φ state information, for example a given observation of the state of the system,

the probability of a state d _± j with a given observation φ with the expected response a _i , and p (di _j / φ) the probability of a state dij with a given observation φ and w _{± j represent} the weight of the variable d ^.

The relevance of information to be obtained is defined on the basis of a function of the changes in the state probabilities of a (sub) set of stochastic variables, the function value of which, with identical probabilities, is minimal with and without information and increases monotonically with increasing difference in the probabilities. The relevance of a question or the gain of information about the system is first evaluated in such a way that the greatest possible changes in knowledge after the question has been asked, even if uncertainty is introduced, increase its relevance. The difference in the probabilities of the presence of a particular state of the system by

is a measure of the gain of an answer after asking a question q _k, when answered a probability of the state d, - is expected. The difference or change in the state probabilities pd _y ) is first normalized to a positive value using an absolute value function abs {x) or by squaring. Thus, the difference between the state probabilities after a question may have been asked can be used as a distance measure. Using the amount function ensures that responses that reduce the likelihood of a condition, for example the vaccination against a certain disease, also make a positive contribution.

The relevance of the question is preferably determined with the regulation

eg (q _k , Φ) = ∑p (α «| φ) e ( ^fl w» ^φ ) t formed where

represents the probability of an answer a _kl given the observation Φ of the system.

After the selection of a question has been determined, further steps which are necessary for completing the inventive classification or diagnosis method are carried out.

Here the expected profit described above is subtracted from the cost of the question. This could be done, for example, with the rule neg (q _k , Φ) = eg (q _k , Φ) -c (q _k ), where c _{k is} the cost of the question, such as the time it takes to get a blood sample, shown. This difference gives the net profit when a specific question is asked. The best question can therefore be asked with q _{k * {φ)} , whereby the condition k * (Φ) = argmax (neg (q _k , Φ)) is preferably set. The question is optimally only asked as soon as the expected net profit of the question is greater than zero. Since a large list of questions would be available at the beginning of a question sequence, which has all the questions about all possible states of a system, it is preferred to sort the list of questions at a given level of information so that a smaller, there is a more targeted list. The information level can be a priori, stochastic distribution, status information or an observation φ of the status of the system. The list of questions is thus influenced by the current level of information in such a way that it excludes certain questions.

With regard to a general stochastic model of a system, the evaluation of the relevance of the questions can also be understood as the evaluation of missing information of the stochastic model. The missing information is the expected answer to a question to be asked. However, it is not necessarily restricted to this form.

The exemplary embodiment of the invention presented here can be supplemented by means of a network, the selected questions being sent to a client with the highest relevance and the answers being received, stored and evaluated accordingly by a server. This guarantees an online expert system for diagnosing the status of a system.

In addition, the expert system according to the invention can determine the actual state of a process, for example a manufacturing process, on the basis of the electronic query of the variables generated in the process by sensors. The missing information of the stochastic model here are the unknown variables of the industrial process, such as temperature of a supply means, pressure, speed of an assembly line, number of certain defects of a product on an assembly line. These states can all of suitable sensors are queried and the procurement of various information by the sensors is controlled by the method described for the ¹ medical method. The results of the evaluation or classification can be used to control the process using a feedback loop. The information recorded by the sensors can be queried automatically by a suitably equipped computer.

It is advantageous that the interactive property of a conventional medical procedure is eliminated here, since an extremely fast reaction time is usually also required for industrial processes. Furthermore, in an emergency, rapid convergence to determine the actual state of the system can be guaranteed in order to strengthen the security and efficiency of the system. The questions and answers as such of the medical procedure would thus run automatically without the intervention of a person.

The following publications are cited in this document:

[1] Gonzalez & Woods: Digital Image Processing pgs. 413 -

458, Addison Wesley, 1993

[2] Gonzalez & Woods: Digital Image Processing, pgs. 571 -

597, Addison Wesley, 1993

[3] F. Jensen: An Introduction to Bayesian Networks

UCL Press, 1996.

Claims

claims

1. Expert system for evaluating a gain in information about a state of a system, comprising: a means for determining a state information of the system,

a control unit for controlling the agent,

- An evaluation unit for evaluating the status information, wherein

- using the agent

- Information (φ 1) is determined about an actual state of a system and

- at least one additional piece of information (φ2) about the actual state of the system is determined,

a first predicted state (Prl (φl)) of the system is determined from the information using the evaluation unit,

a first probability (Pl (St)), with which the first predicted state matches the actual state of the system, is determined, using the combination of the information and the additional information, a second predicted state of the system (Pr2 (a _ki U φ) ) is determined

- A second probability (P2 (St)) is determined, with which the second predicted state matches the actual state of the system, - A difference (D _p ι _p2 ) between the first and the second probabilities is determined and this as an evaluation measure is used for gaining information about the system through the additional information, and - using the control unit, the means are controlled as a function of the evaluation measure.

2. Expert system according to claim 1, wherein the evaluation and control units are part of a computer and are controlled by a computer program product.

3. Expert system according to one of claims 1 or 2, in which the means for determining a status information are sensors.

4. Expert system according to one of the preceding claims, in which the evaluation unit has a knowledge base which has modeled states of the system, and the knowledge base is held by a storage medium.

5. Expert system according to one of the preceding claims, in which the determination of the state information on the

System can be implemented by means of a request for status information about the system.

6. Expert system according to one of the preceding claims, wherein the amount of information gained is a criterion for the priority of a request for additional information about the system.

7. A method for evaluating a gain in information about a state of a system in which

- Using a means for determining a status information of the system

- Information (φ 1) about an actual state of a system is determined and - At least additional information (φ2) about the actual state of the system is determined using an evaluation unit

- a first predicted state (Prl (φl)) of the system is determined from the information, - a first probability (Pl (St)) is determined with which the first predicted state corresponds to the actual state of the system,

a second predicted state of the system (Pr2 (a _k χU φ)) is determined using the combination of the information and the additional information,

- A second probability (P2 (St)) is determined, with which the second predicted state matches the actual state of the system, - A difference (D _p ι _p2 ) between the first and the second probabilities is determined and this as an evaluation measure is used for gaining information about the system by means of the additional information, - the means is controlled on the basis of a control unit as a function of the evaluation measure.

8. The method according to claim 7, in which the evaluation and control units are controlled by a computer program product.

9. Expert system according to one of claims 7 or 8, in which sensors are used as means for determining status information.

10. The method according to any one of claims 7 to 9, in which a knowledge base is used for evaluating state information, which has modeled states of the system.

11. The method according to any one of claims 7 to 10, wherein the determination of status information about the system is carried out by means of a request for status information about the system.

12. The method according to any one of claims 7 to 12, in which as a criterion for the priority of a request for a ^S tandsinformation the amount of information gain is used.