Predicting temporal development of process with improved consideration of prior knowledge about process
Through the collaboration of probability models and process models, combined with historical records and priority knowledge, select the most likely parameter time to develop candidates, solve the problems of insufficient accuracy and multiple long -term accuracy challenges in existing technologies, realized More efficient and accurate predictions.
Patent Information
- Application Number
- CN202411020010.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2023-07-27
- Filing Date
- 2024-07-29
- Publication Date
- 2025-01-28
AI Technical Summary
Prior arts are difficult to effectively utilize prior knowledge when predicting the development of parameter time affected by a specified process, resulting in insufficient prediction accuracy, especially the challenge of maintaining accuracy over multiple time steps.
By using the collaborative work of the probability model and the process model, the historical record VT provides the conditional probability distribution of the parameter X, and combines the priority knowledge to determine the candidate and its scores to select the most likely future time change process.
It has significantly reduced the computing workload, improved the accuracy of forecasting, ensured the compatibility of prediction results and priority knowledge and boundary conditions, and reduced the probability of being missed due to early refusal candidates and then turned to the event.
Smart Images
Figure CN119378675A_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the prediction of the temporal development of a variable which is influenced by a specified process, wherein any a priori knowledge about the process exists and is utilized. Background Art
[0002] In order to predict the temporal development of variables relevant to technological processes, data-based probabilistic models are often used. Such models infer future course of behavior from the previous course of behavior of the variable. In particular, these models indicate how probable certain future values of the variable are, given the previous course of behavior.
[0003] The relationship between previous and future course of change is determined by the mechanisms of action of the examined process. If the model has learned to predict the future development of the variable of interest based on the training time series of this variable, then the model has at least implicitly learned these mechanisms.
[0004] The behavior of many technological processes follows previously known laws and / or is limited by hard boundary conditions. Therefore, only predictions of time developments that are sufficiently compatible with these previously known laws and / or boundary conditions are suitable for practical use. Time developments that seriously violate this cannot be realized in practice. Summary of the invention
[0005] The invention provides a method for predicting the temporal development of a variable x affected by a given process. The method uses an existing probability model based on a history V of time steps t. t To provide the value x of the parameter x t+1 The conditional probability distribution p(x t+1 |V t ).
[0006] The history V t In particular, it may include, for example, the value x of the variable x for the time step t. t , and / or any previous value x of the parameter x at the previous time step tk t-k This means in particular that the probabilistic model does not only have to make predictions based on a single previous value, but can also take into account previous changes over time. Thus, for example, the future steering angle of the vehicle can only be predicted with a relatively high degree of uncertainty based on the current value. However, if the recent history is also taken into account, it can be recognized, for example, that the vehicle is driving around a bend. The prediction of the future value of the variable x can then be, for example, a continuous continuation of the driving around this bend.
[0007] Alternatively or also in combination therewith, the history record V tIt may also include: any other context information C of time step t and / or any previous time step tk t and / or C t-k These contextual information C t and / or C t-k It may be time-dependent or time-independent. For example, when monitoring an industrial process in an industrial facility, the time-dependent context information C t and / or C t-k The parameters generated in the previous process step can be involved. This history provides meaningful changes in the future of the measured variable x that characterizes the current process step, and thus improves the prediction accuracy. The understanding behind this is that with this industrial process, a unified overall goal is tracked as a whole, and the individual process steps cooperate rationally to achieve this goal. Time-independent context information C t and / or C t-k For example, it may relate to the type of process, to the characteristics of the machine and / or the installation and / or to the product currently being manufactured. t and / or C t-k For example, operating conditions may also be involved, such as the temperature at the location of the facility, the air pressure, the air humidity or also the energy price. For example, the age, operating hours and / or degree of wear of the machine or facility may also be taken into account. For example, the number of hours that the current shift of the operating personnel has lasted may also be taken into account. This number of hours may, for example, influence the reaction time of the personnel to a specific event.
[0008] In the scope of this method, based on the history V of time step t t , determine the value x of the parameter x at time step t+1 t+1 k candidates If the measured variable x is a continuous variable, then these candidates For example, it is possible to sample from intervals in which the measured variable can even be moved. For this sampling, for example, an equidistant grid can be used. On the other hand, if the measured variable x is a discrete variable, then the candidate The choice of is restricted from the outset to the possible values that the discrete variable can take. Thus, for a discrete variable x, the value x t+1 The conditional probability distribution p(x t+1 |x t ) is a categorical distribution.
[0009] For the identified candidates The score is determined by combining a probabilistic model with a process model that represents prior knowledge about the specified process. These scores are the corresponding candidates corresponds to the actual value x of the parameter x at time step t+1 t+1In this case, the conclusions of the probability model can be calculated in any way using the conclusions of the process model. For example, as mentioned at the beginning, if the process follows previously known laws and / or is subject to boundary conditions, then the candidates that violate these laws or boundary conditions can be reduced, for example by additive or multiplicative "penalties" Score In extreme cases, the violation itself may result in: is set to the worst possible value (e.g., to 0). This is similar to the fact that when conducting commodity inspection on electrical equipment, a device that works well even when the housing is under power supply voltage will not exceed the "fail" range.
[0010] From the candidate In the set of To select a proper subset The newly selected candidate based on time step t+1 Determine the value x of the parameter x at time step t+2 t+2 l candidates
[0011] It has been recognized that in this way, prior knowledge about the specified process can be taken into account without having to resort to an already existing data-based probabilistic model. As a result, the performance of the probabilistic model does not deteriorate, as might otherwise occur when optimizing for a new optimization objective.
[0012] At the same time, as the above examples show, there is maximum flexibility in how prior knowledge should be considered in detail.
[0013] The probability model and the process model with prior knowledge cooperate synergistically, so that the computational workload for finding the future time variation process that is most likely to correspond to the actual variation process of the parameter of interest within multiple time steps is significantly reduced.
[0014] Based on historical records t , such as the value x of the variable of interest x at time step t t , there are multiple candidates for the value of parameter x at the next time step t+1 To determine the most likely accurate prediction for time step t+2, we must A new candidate for the value of the parameter x at time step t+2 is determined Therefore, the entire tree of candidate time-varying processes for the variable of interest x should be examined. The number of candidate time-varying processes in the tree grows exponentially with the number of time steps over which the prediction extends.
[0015] In this case, a lot of work can be saved if as many candidates as possible are picked at the earliest possible level of the tree and these candidates are not examined at subsequent time steps. This is roughly equivalent to the fact that in the early stages of an epidemic, the cautious behavior of one person who infects one person instead of three (who then do not infect three others, and so on) can have a surprisingly large effect on the total number of cases.
[0016] This is at the expense of possibly missing the global optimum. In principle, it is possible that the first time step of the time development obtains only a moderate score from the probability model, but is followed by the next time steps with very good scores. At the end of the forecast horizon of, for example, four or five time steps, the moderate score of the first time step has long been "forgotten", and the overall score of this time development is still better than the overall scores of other time developments that started out very promisingly in the first time step and later disappointed.
[0017] Therefore, in order to reduce the probability of missing out on a later turn to a still very good time development due to early rejection of a candidate, it is advantageous to preferably reject such candidates for which further examination, in addition to the simple probability score, still contradicts other valid reasons. Here, the process model plays a role: in particular, for example, the candidate whose value of the variable x at the time step t+1 can be identified These candidates certainly do not appear in this form, since they represent states of the process that cannot be reached, for example, due to the known mechanism of operation of the process or due to hard boundary conditions. If the law of conservation of energy, the laws of thermodynamics, or other mandatory physical requirements are not met, then the candidate Other developments of the time are invalid from the outset, because there is no way to get there. This is roughly equivalent to: it is useless for a climber in distress to philosophize about how to immediately deal with their situation using a specific piece of equipment - which they unfortunately left at home. The climber can immediately shelve all plans that require the presence of exactly that piece of equipment, because discussions about this would only take time, which they do not have. Whoever in this situation focuses his energy on alternatives that are feasible with the existing equipment significantly increases his chances of survival.
[0018] In a particularly advantageous embodiment, the process model therefore contains boundary conditions which the temporal development of the variable x and / or the behavior of the process derived therefrom must satisfy. Candidates which can never be realized due to violation of hard boundary conditions are rejected. There is nothing wrong with people.
[0019] Examples of particularly hard boundary conditions that can be used in this regard are:
[0020] Mechanical or electrical constraints; and / or
[0021] Satisfy physical conservation laws; and / or
[0022] Satisfy the physical continuity equation.
[0023] In a particularly advantageous embodiment, the fraction Contains the product of a first contribution from the probability model and a second contribution from the process model. The two contributions can be weighted against each other, for example, by a constant factor. In particular, a contribution of 1 from the process model can indicate that all boundary conditions are met and the process will develop as described by the probability model. On the other hand, candidates that cannot be realized The process model can be excluded from further consideration due to its contribution of 0. Thus, advantageously, the contribution of the process model is 1 when the boundary conditions are fully met, and is 0 when at least one boundary condition is clearly not met.
[0024] Based on score To the candidates A proper subset of The selection can be made based on any suitable criteria.
[0025] Thus, for example, the The top n candidates Included in the proper subset By considering more than one candidate, it is possible in particular to obtain the possibility that a time course which was not initially at the top can be improved afterwards.
[0026] Alternatively or in combination therewith, the fraction Those candidates that exceed the specified threshold Included in the proper subset In this way, it is possible to ensure that: In the set of is not automatically selected as the best candidate. More precisely, it is likely that: is empty. This can then be evaluated as an error signal, so that the actual behavior of the process can no longer be described by the probabilistic model. This can prompt the operator of the process to look for errors, such as equipment defects.
[0027] In a further advantageous embodiment, a process model is selected which describes the behavior of the process using at least one differential equation. The differential equation can be used, for example, in particular to predict the time behavior of the process under certain simplifying assumptions. Thus, candidates that agree with the behavior predicted by the differential equation are than candidates that cannot be explained by differential equations at all is more likely to reflect the actual expected behavior of the process.
[0028] For example, at least one of the differential equations can be an ordinary differential equation. Then, the process model is a fraction The contribution of the candidate can be measured: The degree of deviation from the prediction determined using the ordinary differential equation. In this case, whenever the probability model is trained on real data of the process, there will be some deviation because the differential equation is a simplified description of the process.
[0029] For example, at least one of the differential equations may also be a stochastic differential equation. The stochastic differential equation provides a probability distribution for the expected future values of the variable x of interest. Then, the process model is The contribution of can be measured: According to such a probability distribution, the candidate What is the probability of . Therefore, this contribution does not require any reinterpretation from deviation to probability, but exists directly as a probability.
[0030] In another particularly advantageous embodiment, the candidate and the relevant scores Listed in Based on Score On a sorted list, where n ≥ 0. From this list, select the one with the best score Candidates Based on the candidate Determine the value x of the parameter x at time step t+n+1 t+n+1 New Candidates In this case, the list is likely to contain candidates involving different future time steps t+n Next, always target the following candidate To continue working, for this candidate, according to the current state of knowledge, this candidate is most likely to correspond to the actual value x of the parameter x at time step t+n t+n . Then, for example, it might happen that candidate x 1 Initially leading the list in time steps t+1 and t+2, but then only trained with all the possibilities with very bad scores in time step t+3. Then, all these possibilities may lag behind, for example, candidate x which was ranked second in the list so far by development up to time step t+1.2 As a result, candidate x 2 The time development of then advances to time step t+2 and, when score then still remains at the top of the list, to time step t+3.
[0031] In another particularly advantageous embodiment, each time a candidate is determined Score Afterwards, the candidate Then, from the updated list, select the one with the best score Candidates Then, before checking the first further development of the candidate from time step t+2 to time step t+3, in particular, for example, it is not necessary to first check all possibilities of how time step t+2 can be reached from time step t+1. Rather, as long as the first possibility is promising, it can be directly continued to be checked up to the later time steps t+3, t+4, and so on.
[0032] In another particularly advantageous embodiment, the list has a specified capacity, when the specified capacity is exceeded, the list with the worst score is rejected. The candidate By specifying the capacity, the search strategy can be set. If the list has an infinite capacity, the search strategy is transformed into a full search of the possibility tree, and its workload grows exponentially with the number of time steps. If the list has a capacity of only one candidate, the search strategy is transformed into a "greedy algorithm", which always searches for the best option only in the corresponding next time step, so that the time step is then considered to be completed and immediately turns to the next time step. In the case where the capacity is between 1 and infinity, any intermediate form between the "greedy algorithm" and the full search can be set. This intermediate form is also called "beam search" (English: beamsearch). The beam search checks the graph in the following way: the beam search selects the most promising nodes from the available number of nodes in each step and further expands these nodes. Therefore, compared with the search using the "greedy algorithm", the beam search takes into account the entire predicted time series.
[0033] The limited size of the list ensures that, on the one hand, the heuristic search is faster than a complete search of the possibility tree, and, on the other hand, the memory requirements are not excessive. As mentioned above, this comes at the expense of the fact that the temporal evolution of the final prediction of the parameter x is not necessarily the most accurate prediction in practice. It is no longer guaranteed that the global optimum is not hidden in an earlier candidate that was rejected at time step t+1. However, by taking into account the prior knowledge in the process model, the probability of this happening is reduced. In specific applications, the best results are usually not required, but practical results are desired within a limited time and with limited hardware resources.
[0034] In a further particularly advantageous embodiment, an actuation signal is determined as a function of the determined temporal development of the variable x. The vehicle, driver assistance system, robot, power tool, domestic appliance and / or medical diagnostic device is actuated with the aid of the actuation signal. In this way, the probability that the reaction executed by the respectively controlled system is adapted to the situation reflected in the temporal development of the variable x is increased.
[0035] For example, in vehicle applications, the time development of the speed or trajectory of a vehicle can be predicted. The predicted time development of the speed can be used, for example, to simulate the exhaust behavior of vehicles in order to determine whether the pollutant emissions of these vehicles under real driving conditions meet legal requirements. In particular, it can be checked how the exhaust behavior depends on the engine control parameters, so that vehicles seeking new model approval can pass the exhaust test under real conditions as much as possible at the first attempt. The predicted trajectory of one's own vehicle and other traffic participants can be used, for example, to make the at least partial automatic driving of the vehicle more forward-looking. If, for example, it can be foreseen that the vehicle must slow down or even stop due to traffic, the engine can be switched to generator operation for electrically driven vehicles in order to store the kinetic energy at least partially in the battery. The predicted trajectory can also be used to determine the total load of the component and, for example, to initiate state-dependent maintenance. Similarly, for power tools or household appliances, the wear state can also be predicted as the variable of interest x and the corresponding state-dependent maintenance can be initiated.
[0036] In human-controlled vehicles, for example, the driver's behavior can also be predicted as a variable of interest x. For this purpose, multiple data sources can be used, such as video recordings of eye movements, video recordings of steering movements, or health data such as heart rate determined by the driver's smart watch. For example, the temporal development of the driver's fatigue or exhaustion state can be predicted so that corresponding warnings can be output more accurately. If it is foreseeable that the driver will still reach his destination without any problems, there is no point in issuing warnings to him along the way. These warnings may eventually become a self-fulfilling prophecy and make the driver even more inattentive. On the other hand, if it is foreseeable that the destination can no longer be reached safely, it makes sense to bring this to the driver's attention as early as possible so that he can try to find alternatives.
[0037] In the field of smart homes and industrial buildings, for example, the concentration of oxygen or carbon dioxide, air humidity, atmospheric pollution or temperature can be predicted as the variable x of interest. In this way, measures can be proactively taken to improve the indoor climate. For example, energy-saving floor heating has a high thermal inertia, so that it should be activated as early as possible when needed so that the indoor temperature does not drop too much - which in turn may induce residents to use electric heaters, which would impair the energy efficiency of the floor heating. For example, in the case of an expected significant increase in the temperature of a room, the air conditioning system can also be proactively switched on when electricity is particularly cheap.
[0038] In all of these and other applications, promising candidates for the next time step t+1 can be improved by taking into account the corresponding relevant prior knowledge in the process model as described here. The selection of
[0039] The method can in particular be completely or partially computer-implemented. Thus, the invention also relates to a computer program having machine-readable instructions which, when executed on one or more computers and / or computing instances, cause the one or more computers to perform the described method. In this context, control devices for vehicles and embedded systems for technical devices which are likewise capable of executing machine-readable instructions are also to be regarded as computers. Computing instances are, for example, virtual machines, containers or other execution environments in which computing power is provided in the cloud.
[0040] Likewise, the invention also relates to a machine-readable data carrier and / or a download product having the computer program. A download product is a digital product that can be transmitted via a data network, that is to say can be downloaded by a user of the data network, which digital product can be sold for immediate downloading in an online store, for example.
[0041] Furthermore, one or more computers and / or computing instances can be equipped with the computer program, the machine-readable data carrier or the downloaded product. BRIEF DESCRIPTION OF THE DRAWINGS
[0042] Further measures for improving the invention are presented in more detail below together with the description of preferred exemplary embodiments of the invention based on the drawings.
[0043] in:
[0044] Figure 1 An embodiment of a method 100 for predicting the temporal development of a variable x is shown;
[0045] Figure 2 Shown according to the score Candidates for sorting An exemplary chronological development of the list. DETAILED DESCRIPTION
[0046] Figure 1 is a schematic flow chart of an embodiment of a method 100 for predicting the temporal development of a variable x affected by a given process. For this prediction, a probability model 1 is used, which, given a history V up to a time step t, t Provides the value of parameter x in the case of t+1 The conditional probability distribution p(x t+1 |V t ). On the other hand, a process model 2 is also used, which represents the prior knowledge about the specified process.
[0047] In step 110, based on the historical record V of time step t t , determine the value x of the parameter x at time step t+1 t+1 k candidates
[0048] In step 120, for the candidate Determine the score by combining probability model 1 with process model 2 These scores Is a candidate corresponds to the actual value x of the parameter x at time step t+1 t+1 A measure of the probability of .
[0049] In step 130, from the candidate In the set of To select a proper subset
[0050] In step 140, the newly selected candidate based on time step t+1 Determine the value x of the parameter x at time step t+2 t+2 l candidates Next, these l candidates The process can return to step 120. That is, for the l candidates Then you can redefine the score And can select promising candidates A subset of is used for continued tracking in time step t+3.
[0051] In this way, from the history V at time step t t , i.e., for example, based on the value x of the measured variable x at time step t t, the prediction x of the measured variable x within the range of h time steps in the future can be determined continuously t+1 ,…,x x+h .
[0052] exist Figure 1 In the example shown in FIG. 1 , in step 150, the predicted x is determined from the t+1 ,…,x x+h In step 160, the vehicle 50, the driving assistance system 51, the robot 60, the power tool 70, the household appliance 80 and / or the medical diagnostic device 90 are controlled by using the control signal 150a.
[0053] According to block 121, the process model 2 may contain boundary conditions, which the time development of the variable x and / or the behavior of the process resulting from this time development must satisfy. According to block 121a, these boundary conditions may include, for example, in particular:
[0054] Mechanical or electrical constraints; and / or
[0055] Satisfy physical conservation laws; and / or
[0056] Satisfy the physical continuity equation.
[0057] According to block 122, the score It may contain the product of a first contribution from probability model 1 and a second contribution from process model 2.
[0058] In particular, according to block 123, when the boundary conditions are met, the contribution of the process model may be 1, and when at least one boundary condition is not met, the contribution of the process model may be 0. As described above, by doing so, the candidates that do not meet these boundary conditions may be are completely excluded from further examination, as these candidates are anyway unrealizable in practice.
[0059] According to block 124, a process model 2 may be selected that describes the behavior of the process using at least one differential equation.
[0060] In this case, in particular, for example, according to block 124a, at least one differential equation may be an ordinary differential equation. Next, according to block 124b, the process model 2 calculates the fraction The contribution of the candidate can be measured: The degree of deviation from the prediction determined using ordinary differential equations.
[0061] Alternatively or in combination therewith, according to block 124c, at least one differential equation may be a stochastic differential equation. Next, according to block 124d, the process model (2) is used to calculate the fraction The contribution of can be measured as follows: According to the probability distribution determined by using stochastic differential equations, the candidate What is the probability.
[0062] According to block 131,
[0063] · Have the best score The top n candidates and / or
[0064] Its score Those candidates that exceed the specified threshold
[0065] Included in the proper subset of candidates to be considered for the next time step middle.
[0066] According to block 132, the candidate and the relevant scores Can be listed based on score The sorted list, where n ≥ 0. Then, according to block 133, the one with the best score can be selected from them. Candidates Next, according to block 141, based on the candidate The value x of the parameter x at time step t+n+1 can be determined t+n+1 New Candidates
[0067] In particular, according to block 132a, each time a candidate is determined Score Afterwards, the candidate can be Then, according to block 133a, from the list thus updated, the candidate with the best score can be selected. Candidates
[0068] Figure 2 The example illustrates: using the score Candidates for sorting A list of how promising predictions for the variable of interest x can be determined over multiple time steps.
[0069] The forecast is based on the historical record V at time step t t , such as the value x of a parameter x at time step t t From this, a number of candidates are determined for the value of the parameter x at the first future time step t+1, of which only two are plotted for clarity. and exist Figure 2 In the example shown in In the sorted list, candidates First place, and the candidate Ranked second.
[0070] To continue the prediction until time step t+2, just keep tracking the candidate That is, for the candidate Compute other candidates for time step t+2. Candidates Remain unchanged. Figure 2 In the example shown in The score of is so good that this candidate is ahead in the list over candidates computed only up to time step t+1
[0071] But if the most promising candidate Continuing to time step t+3, there is an "unfortunate surprise": all candidates generated in this way have bad scores. The score is still the best, but it is also better than the previous second-ranked candidate. are only calculated until the score at time step t+1 is worse. Thus, there is now a change of leadership in the list: Candidate No. 1 candidate Only ranked 2nd.
[0072] In the next step, this leads to: Candidate Now trained until time step t+2. Candidate Remain unchanged. The candidate is in the new candidate The second place in the following ranking list.
[0073] Therefore, the heuristic search strategy in this example is better than the greedy algorithm, which does not take into account the dependencies between time steps and thus ends up rejecting candidates early.
Claims
1. A method (100) for predicting the temporal evolution of a variable x affected by a specified process using a specified probability model (1), said probability model being given a history V up to a time step t. t Provides the value of the parameter x in the case t+1 The conditional probability distribution p(x t+1 |V t ), the method comprising the following steps: Based on the history V of the time step t t , determine (110) the value x of the parameter x at the time step t+1 t+1 k candidates For candidates Determine (120) a score by combining the probability model (1) with a process model (2) representing prior knowledge about the specified process The scores are the corresponding candidates corresponds to the actual value x of the parameter x at the time step t+1 t+1 A measure of the probability of From the candidate In the set of To select (130) proper subset The newly selected candidate based on the time step t+1 Determine (140) the value x of the parameter x at the time step t+2 t+2 l candidates 2. The method (100) according to claim 1, wherein: The process model (2) contains (121) boundary conditions, which the time evolution of the variable x and / or the behavior of the process resulting from the time evolution must satisfy.
3. The method according to claim 2, wherein: At least one boundary condition contains (121a) Mechanical or electrical constraints; and / or Satisfy physical conservation laws; and / or Satisfy the physical continuity equation.
4. The method (100) according to any one of claims 1 to 3, wherein: The score It comprises (122) the product of a first contribution of the probability model (1) and a second contribution of the process model (2).
5. The method (100) according to any one of claims 2 to 3 and additionally claim 4, wherein: When the boundary conditions are met, the contribution of the process model is (123) 1, and when at least one boundary condition is not met, the contribution of the process model is 0.
6. The method (100) according to any one of claims 1 to 5, wherein · Have the best score The top n candidates and / or Its score Those candidates that exceed the specified threshold is included in (131) into the proper subset middle.
7. The method (100) according to any one of claims 1 to 6, wherein: A process model (2) is selected (124), the process model utilizing at least one differential equation to describe the behavior of the process.
8. The method (100) according to claim 7, wherein At least one differential equation is an ordinary differential equation (124a), and The process model (2) has an effect on the score Contribution measurement (124b): Candidate The degree of deviation from the prediction determined using the ordinary differential equation.
9. The method (100) according to any one of claims 7 to 8, wherein At least one differential equation is a (124c) stochastic differential equation, and The process model (2) has an effect on the score Contribution measure: According to the probability distribution determined by the stochastic differential equation, the candidate What is the probability of (124d).
10. The method (100) according to any one of claims 1 to 9, wherein Candidate and related scores Listed (132) in the On a sorted list, where n ≥ 0, · Select (133) from them with the best score Candidates and Based on the candidate Determine (141) the value x of the parameter x at the time step t+n+1 t+n+1 New Candidates 11. The method (100) according to claim 10, wherein · Each time a candidate is identified Score Afterwards, the candidate into said list, and From the list thus updated, select (133a) with the best score Candidates 12. The method (100) according to any one of claims 10 to 11, wherein: The list has a specified capacity, and when the specified capacity is exceeded, the list with the worst score is rejected. The candidate 13. The method (100) according to any one of claims 1 to 12, wherein determining (150) a control signal (150a) as a function of the determined temporal development of the variable x; and Using the control signal (150a), a vehicle (50), a driving assistance system (51), a robot (60), an electric tool (70), a household appliance (80) and / or a medical diagnostic device (90) is controlled (160).
14. The method (100) according to any one of claims 1 to 13, wherein: The historical record V t include: The value x of the parameter x at the time step t t ; and / or The previous value x in time of the parameter x at the previous time step tk, k>0 t-k ; and / or Context C at time step t and / or any previous time step tk, k>0 t and / or C t-k .
15. A computer program comprising machine-readable instructions which, when executed on one or more computers and / or computing instances, cause the one or more computers and / or computing instances to perform a method (100) according to any one of claims 1 to 14.
16. A machine-readable data carrier and / or a download product having a computer program according to claim 15.
17. One or more computers and / or computing instances having a computer program according to claim 15 and / or having a machine-readable data carrier and / or a download product according to claim 16.