Machine parameter optimization using random correction

Abstract

1. A method for optimizing parameters of a machine, comprising the steps of: randomly determining correction values ​​for correcting current values ​​of the parameters based on their respective current values ​​and a step size; modifying the parameters at the machine to the corrected values; evaluating an output of the machine achieved using the corrected parameters; fitting a linear function model of the parameters based on the evaluated output; estimating an effect on the output if the parameters are modified by at least the step size using the linear function model; taking into account the estimated effect and deciding whether to modify the parameters based on a desired output; and if so, modifying the parameters by at least the step size.

Description

Technical Field

[0001] (Cross-Reference to Related Applications) This application claims priority to European Patent Application No. 21190945.2 filed on August 12, 2021, the entire contents of which are incorporated herein by reference.

[0002] (Field of the Invention) The present disclosure relates generally to machine parameter optimization. Certain aspects relate to a method for optimizing at least one parameter of a production machine, a computer program for performing such a method, a computer program product storing such a computer program, a computer apparatus comprising such a computer program product, and a production machine operably coupled to such a computer apparatus.

Background Art

[0003] Production machines, for example not only stretch blow machines but also other types of production machines, typically have a large number of parameters that can be set to certain settings in order to control the operation of the production machine. In this way, the machine can generate product output, but can also produce other types of output, which can be interpreted as output signals of the production machine, such as noise, heat, electromagnetic radiation, station alignment, machine durability or other maintenance-related signals, consumption of energy (electricity and / or fuel), and / or consumption of other resources such as water and raw materials.

[0004] Typically, a human operator sets these parameters to achieve the desired output. However, this relies on the know-how of a skilled operator. Therefore, problems can arise when an inexperienced operator, who may not be familiar with the relationship between parameter settings and the desired output, operates the production machine. Furthermore, in many production facilities, the production machine operates for extended periods, for example, across multiple operator shifts. This means that a new operator or group of operators becomes responsible for a new shift, but may not be familiar with the machine's condition during the previous shift, and therefore may not be able to operate the machine properly.

[0005] U.S. Patent No. 7,933,679 (B1) discloses the analysis and optimization of a machine process by combining results from finite element analysis, mechanism modeling, and vibration analysis (independently or in combination). The overall goal of the analysis and optimization is to determine specific parameters that provide a high material removal rate while maintaining reasonable tool life (since machining includes milling, drilling, turning, and boring). The cited simulation methods are used to establish tolerances for specific process embodiments. Machining parameters with results within these tolerances are used to update the machine's numerical control (NC) program. After the NC is updated, it is validated by factory testing in the machine shop to detect problems related to tool life or part quality. [Overview of the Initiative]

[0006] In recent years, more and more production machines are equipped with sensors, and their parameters are set via computers.

[0007] The inventors have identified opportunities to better formulate and systematize the methods by which operators operate such production machines by using novel functions of sensors and computer setting parameters.

[0008] Therefore, an object of the embodiments of this disclosure is to address this opportunity. A further object of the embodiments of this disclosure is to overcome any of the problems described above. Another object of the embodiments of this disclosure is to improve the operation of production machinery in general. Furthermore, another object of the embodiments of the invention is to do so online, i.e., during the operation of production machinery.

[0009] In the first aspect, a method for optimizing at least one parameter of a production machine, - A step of randomly determining at least one correction value for correcting the current value of at least one parameter, which is based on the current value of at least one parameter and a predefined step size for at least one parameter. - In a production machine, the step of modifying at least one parameter to its respective modified value, - A step of evaluating the output of a production machine, which is achieved using at least one modified parameter, - A step of fitting a linear function model for at least one parameter based on the evaluated output, - A step of estimating at least one effect on the output of the production machine using a linear function model, given that each parameter is modified by at least its respective predefined step size, - In a production machine, taking into account at least one estimated effect, a step of determining whether to modify at least one parameter by at least its respective predefined step size, based on the desired output of the production machine, -If it is decided to modify at least one parameter, a method is provided which includes the step of modifying at least one parameter by at least its respective predefined step size.

[0010] By randomly determining at least one modification value and modifying at least one parameter accordingly, and by evaluating the output and considering the estimated effect of the potential modification of at least one parameter, this method enables noisy optimization, and the random changes allow the method to effectively learn about the effects of changing parameters while still maintaining safe and predictable behavior. Due to this noisy optimization, the method can more easily discover good modifications than if it followed a fixed, static, pre-set plan. Furthermore, due to the use of a linear function model to determine the effects, the results of modifying at least one parameter can be more easily determined due to the relatively simple nature of the model. Moreover, it can be usefully assumed that small local changes in at least one parameter may lead to small local predictable changes in the output, and even if larger changes in at least one parameter may lead to unpredictable changes in the output, this can be efficiently modeled using a linear function model.

[0011] In a more advanced embodiment, at least one correction value is randomly determined within a range from a value at least one step size lower than the current value of each parameter to a value at least one step size higher than the current value of each parameter.

[0012] In this way, the method can use an efficient range for determining random values. Furthermore, advantageously, in this manner, it can be assumed that at least one modified value is very close to its respective current value and can therefore be easily used to modify its respective current value without incurring a significant risk of disrupting the operation of the production machine.

[0013] In a more advanced embodiment, the method includes repeating the steps of the method over multiple iterations.

[0014] In this way, the optimal state can be reached over time.

[0015] In a more advanced embodiment, the linear function model is fitted to at least one parameter based on at least one previously evaluated output and at least one previous setting of at least one parameter of the production machine corresponding to at least one previously evaluated output.

[0016] In this way, this method can take history output into consideration.

[0017] In a more advanced embodiment, the sample importance weights of at least one previously evaluated output and at least one previous setting decrease over time according to a time decay function, preferably an exponential decay function.

[0018] In this way, the method can take into account the physical and commercial reality that recent output exceeds historical output. In addition, this makes it possible to consider the time course of optimization of production machinery in order to give improved weight to various production factors in order to improve optimization.

[0019] In a more advanced embodiment, at least one effect is estimated by approximating at least one derivative of at least one parameter and determining at least one value of the at least one approximated derivative when each parameter is modified by at least its respective predefined step size.

[0020] In a more advanced embodiment, at least one parameter is multiple parameters, and, based on the more advanced embodiment described above, at least one approximate derivative is the gradient.

[0021] In this way, the method allows us to interpret the gradient to determine where it leads to a better output, and therefore determine which parameter(s) should be modified and in what way.

[0022] In a more advanced embodiment, the method includes the step of performing a heuristic configured to determine the optimal combination of modifications to at least one parameter to approach a desired output, taking into account at least one estimated effect.

[0023] In this way, this method can efficiently select the appropriate direction when the heuristic has many parameters, thus leading to a focus on the parameter that is likely to have the greatest influence. When there are only a few parameters, a brute-force evaluation of all possible parameter combinations remains computationally manageable and can lead to the optimal direction.

[0024] In a more advanced embodiment, determining whether or not to modify at least one parameter involves minimizing or maximizing a loss function, which is preferably based on a sigmoid function and / or the sum over time of the differences between the evaluated output and the desired output.

[0025] In this way, the loss function may be better explainable; that is, by using an easily understandable mathematical function, the loss function may be designed so that its operation and circumstances are more clearly understood by humans. Furthermore, the sigmoid function and the sum over time may enable more efficient operation of this method, as both functions enable high-speed computation.

[0026] In a more advanced embodiment, if the steps of the method are repeated over multiple iterations, the predefined step size of each step changes at least once over one of the multiple iterations, preferably decreasing.

[0027] In this way, the method can incorporate a measure of flexibility in the iterative process. This is advantageous because it may initially optimize more quickly by approaching the minimum step size, but later optimize more accurately and predictably. If the step size is increased over iterations, optimization may proceed more rapidly and / or the optimization may overcome local but non-global optimal conditions, which can be advantageous, for example, if an optimal condition is found that actually yields an unsatisfactory output.

[0028] In a more advanced embodiment, each predefined step size is determined for at least one parameter based on the minimally distinguishable discrete granularity of at least one parameter in the production machine.

[0029] In this way, the step size can correspond to the actual realities of the production machine parameters, for example, allowing only specific changes instead of arbitrary analog changes.

[0030] In a second aspect, a computer program is provided which includes instructions configured to perform the above-described method when executed on a computer processor.

[0031] In a third embodiment, a computer program product is provided that includes a computer-readable medium for storing the aforementioned computer program.

[0032] In a fourth aspect, a computer device is provided which includes the computer program product described above and is configured to perform the method described above.

[0033] In a fifth aspect, a production machine is provided that is operably coupled to the computer device described above.

[0034] Those skilled in the art will understand that the advantages and considerations applicable to the above-described methods also apply to computer programs, computer program products, computer equipment, and production machinery, with the necessary modifications. [Brief explanation of the drawing]

[0035] Aspects of this disclosure will be better understood with the help of the examples described below and the accompanying drawings. [Figure 1] This figure schematically shows a flowchart of one embodiment of the method according to this disclosure. [Figure 2] This figure schematically illustrates a production machine having multiple parameters for optimization using an embodiment of the method of this disclosure. [Figure 3] This diagram schematically shows a graph of the parameters of an optimized production machine. [Figure 4] This figure schematically shows an example of a loss function for use in one aspect of the method disclosed herein, in four graph frames. [Figure 5] This figure schematically shows, in six graph frames, an example of six parameters of a production machine in operation according to one embodiment of the method of the present disclosure. [Figure 6] This figure schematically shows an example of the distribution of plastic material in a bottle during operation according to one embodiment of the method of this disclosure, using four graph frames. [Modes for carrying out the invention]

[0036] In the following description, the term "production machine" can refer to any machine involved in a production process, which produces products that form part of its output, and which can also influence other outputs such as signal outputs, for example, heat or noise generated by the machine (typically as a by-product), or resources such as power, materials, or production time consumed by the machine to produce the product. The embodiments of this disclosure are thought to be applicable to any machine having such outputs. In practical examples, it may be preferable if the waiting time in the production process between modifying parameters and obtaining the effect on production is relatively short.

[0037] In a specific example, the operation of a production machine, such as the stretch blowing process of a stretch blowing machine, can be described as follows:

[0038]

number

[0039] Loss function In various examples, it is preferable to use a loss function. Consider the following example of a loss function: L(Ttop ,T middle ,T bottom ,T base )=l L may represent a loss function designed to convey given instructions to an optimization algorithm. In the specific example of the stretch blow process of a stretch blow machine provided above, such an instruction may be that "more plastic is required at the three sidewall measurement locations and less at the base". The result of the loss function l may be minimized to achieve a more optimal process. Of course, as will be appreciated by those skilled in the art, the mathematical definition of the loss function may be reversed such that the result must be maximized to achieve a more optimal process.

[0040] In the following, T top , T middle , T bottom , T base refers to a vector comprising measurements of specific outputs of a production machine. For example, in the specific example given above, in the context of a stretch blow machine, this may be thickness measurements of a population of bottles. It is considered implicit that some method for measuring or evaluating the output of the production machine is available.

[0041] In various examples, the value of the loss function may be the sum of sigmoid functions that take such a set of thickness measurements as input. A sigmoid designed to instruct an increase in plastic may be of the following form:

[0042] [Formula] On the other hand, the one that instructs a decrease in plastic may be of the following form:

[0043] [Formula] where T is the set of thickness measurements, s is the maximum value of the sigmoid (the minimum value thereof is 0), k is the position of the sigmoid inflection point, and n is the steepness thereof. S p(T) becomes 0 from s as T increases, S N (T) increases from 0 to s as T increases.

[0044] therefore:

[0045]

number

[0046] Here, T S is, T top , T middle , T bottom , T base It is one of them.

[0047] Loss Function Usage Example 1: In the first example, a stretch blow machine producing bottles is in operation, the production process is running, and the operator assumes there is not enough plastic on the sidewalls of the bottles, and that some plastic may be removed at the bottom, at least to some extent. If measurements at all four locations yield an average thickness of, for example, 5 units, as used in the production process, the loss function can be defined as follows: L(T top ,T middle ,T bottom ,T base )=S P (T top (1, 5, 10)) + S P (T middle (1, 5, 10)) + S P (T bottom (1, 5, 10)) + S P (T top (1, 5, 10)) + S P (T base (1,2,10)) Therefore, we add plastic to all positions on the sidewall and instruct the optimization algorithm to ignore the base to some extent. If all other things are equal, the value of the loss function may begin to increase when the measured reading approaches 2 units at the base. Minimizing this exemplary loss function should result in optimizing the dimensions of the bottle.

[0048] Loss function usage example 2: In the second example, the stretch blow machine is in operation, the production process is running, and the operator believes that the results of the process are satisfactory and should be maintained, for example, the measurements for all positions are in units of 5, as used in the production process. L(T top ,T middle ,T bottom ,T base )=S P (T top (1, 4, 10)) + S P (T middle (1, 4, 10)) + S P (T bottom (1, 4, 10)) + S P (T top (1, 4, 10)) + S P (T base (1, 4, 10)) + S N (T top (1, 6, 10)) + S N (T middle (1, 6, 10)) + S N (T bottom (1, 6, 10)) + S N (T top (1, 6, 10)) + S N (T base (1, 6, 10))) This allows us to define a loss function where the global minimum is an average thickness of 5 units across all sensors. When optimization is being performed, the algorithm may attempt to maintain the current thickness of 5 units even as U changes over time.

[0049] To facilitate operator operation, it is preferable to provide useful shortcut commands such as stabilizing or increasing, which will allow the system to automatically determine which scenario should be executed.

[0050] optimization The problem of optimizing production machinery is a physical problem, and evaluating the value of the loss function for a given set of parameters involves the bottles being manufactured and measured. By utilizing an optimization method with iterative parameter corrections bounded by previous states, the produced bottles will ultimately have two or more holes, thus avoiding potentially undesirable destructive evaluations.

[0051] In one example of the method according to this disclosure, optimization can begin with a set of parameters that result in bottles, and the loss function can be designed based on the improvements made to these bottles.

[0052] Each P active A set of boundaries can be defined, namely the minimum value, the maximum value, and the step size allowed for the method to modify the parameters.

[0053] An example of an embodiment of this method can proceed as follows. 1.P active The parameter value is randomly modified a given number of times (in its current form, if the parameter has a value P at the start of the step and its maximum step size is s, then no matter how many random modifications occur, the parameter value will remain [Pk * s,P+k * This will result in staying within [s], where k is a natural number, preferably 1). 2.P active , P passive , T top , T middle , T bottom , T base Collect data from a specific point in time, from the past to the present. 3. The output is the thickness measurement value, and the input is P active and P passiveUsing the simultaneous values, a linear function model (other suitable linear function models may also be used, e.g., ridge regression) is fitted to each thickness measurement. The sample weights in the training set used to fit the linear function model may preferably decrease exponentially over time. 4. Based on the parameters of the fitted model, and assuming that S is locally linear, P is set accordingly. active The contribution to various bottle thicknesses is calculated by increasing or decreasing the step size. 5. In order to minimize the loss function, that is, to achieve the greatest reduction in the loss function, P active We run heuristics to determine the best combination of modifications. 6.P active Update it, that is, modify the parameters of the production machine.

[0054] This process can be optionally repeated over multiple iterations.

[0055] Note for Step 1: Random parameter correction is performed when the model fitted in Step 3 is adjusted for individual P for bottle thickness. active This can make it possible to capture the contribution of parameters. By using randomly determined modification values ​​for parameter correction, the optimization process becomes, so to speak, noisier, which is advantageous as it allows the optimization to learn more effectively. Furthermore, by somewhat limiting the range in which the modification values ​​can be randomly determined, safe and predictable behavior can still be maintained, which is important in physical production processes.

[0056] Step 3 Note: The ultimate goal of the fitted model is not necessarily to make very accurate predictions, and therefore, a classical training / test protocol may not be necessary. This model can be used to approximate the (possibly multidimensional) derivative of S at the current set of parameters, and the coefficients of the equation describing the fitted model are P to the problem. active This may allow us to extract the contribution of P. active , P passive And it can be expected that all of U will diverge. The older the data points, the less relevant they may be for locally approximating the derivative of S. This aspect can be favorably handled by time-decaying the sample weights during the model fitting stage.

[0057] Figure 1 schematically shows a flowchart of one embodiment of Method 100 according to the present disclosure. Method 100 is for optimizing at least one parameter of a production machine 110 and consists of steps 101 to 107.

[0058] Step 101 involves randomly determining at least one correction value to correct the current value of at least one parameter, the at least one correction value being based on the current value of at least one parameter and the respective predefined step size for at least one parameter.

[0059] Each current value may be read directly, or it may be retrieved from a repository with a current value of 108. Each predefined step size may be retrieved from a repository with such a step size of 109.

[0060] Step 102 includes modifying at least one parameter in the production machine 110 to its respective modified value.

[0061] Step 103 involves evaluating the output 111 of the production machine 110, the output being achieved using at least one modified parameter.

[0062] Step 104 involves fitting a linear function model to at least one parameter based on the evaluated output.

[0063] Step 105 includes estimating at least one effect on the output of the production machine 110 using a linear function model for at least one parameter, provided that each parameter is modified by at least its respective predefined step size.

[0064] Step 106 includes determining whether to modify at least one parameter in the production machine 110 by at least its respective predefined step size, based on the desired output of the production machine 110, taking into account at least one estimated effect.

[0065] Step 107 includes modifying at least one parameter by at least its respective predefined step size, if it is decided to modify at least one parameter.

[0066] In some aspects of the present disclosure, if a production machine has built-in control functions for controlling the operation of the production machine, for example, for controlling the delivery of power to heating elements to achieve a set temperature, it may be advantageous to disable the built-in control functions of the production machine or at least design around them. This is because such built-in control functions may otherwise be interfered with and / or interfered with by the operation of those aspects of the present disclosure. Details of how to implement this disablement or this peripheral design are left to those skilled in the art.

[0067] In practical embodiments of the methods described herein, the methods may be implemented as a computer program comprising, for example, instructions in a programming language configured to execute the methods when executed on a computer processor. The programming language may be any preferred programming language, including high-level programming languages ​​such as C, C++, Java®, C#, Python, or Clojure, or combinations thereof, or machine programming languages.

[0068] It will be apparent to those skilled in the art that a computer device can be provided that includes a computer program product, which includes a computer-readable medium for storing such a computer program, and that such a computer device can be operably coupled to a production machine, for example, by channels for communicating data and control signals. It is implicitly assumed that the computer device includes a computer processor.

[0069] In a more advanced embodiment, the steps of the method may be repeated over multiple iterations. Referring to Figure 1, this more advanced embodiment can be achieved by not stopping the execution of the method after step 107, but rather returning to step 101. In subsequent iterations of the method flow, the current value 108 may have been updated by the previous iteration. Similarly, the production machine 110 is in a different state compared to the previous iteration, and therefore, the output 111 is also most likely to be different compared to the previous iteration.

[0070] Preferably, the method is stopped when a predefined stopping criterion is reached, for example, when the value of the loss function described above falls below a predefined threshold. Preferably, the method may be started or restarted when a predefined starting criterion is reached, for example, when the value of the loss function described above exceeds a predefined threshold.

[0071] Figure 2 schematically shows one embodiment of the method according to this disclosure, for example, a production machine 200 having several parameters 201 to 203 for optimization using the same method 100 as in Figure 1, in which case the production machine 200 functions as the production machine 110 in Figure 1.

[0072] This figure shows that production machine 200 derives output 207 using several parameters of production machine 200, three of which are parameter 1, indicated by reference numeral 201; parameter 2, indicated by reference numeral 202; and parameter N, indicated by reference numeral 203. Naturally, the symbol N indicates that there can be any number of parameters of production machine 200, and is shown as an abbreviation between parameter 2 and parameter N.

[0073] For clarity, some features are shown only for parameter N, but similar features may exist for other parameters, and similar considerations apply to them as well.

[0074] Parameter N is coupled with a predefined step size N, indicated by reference numeral 204. Step size N may be, for example, the smallest or most favorable step size for parameter N. In a particular example, if parameter N represents oven temperature, step size N may be 0.1°C, which is the smallest step size allowed by the production machine 200, or 0.5°C or 1°C, whichever is more favorable for the operator or for optimization according to the embodiments of the method of this disclosure. In a more advanced embodiment, step size N may be increased or decreased throughout the optimization process if it is deemed useful for optimization. For example, in the early stages of optimization, it may be advantageous to use a relatively large step size, e.g., 5°C, as a favorable granularity, while in later stages of optimization, it may be advantageous to use a relatively small step size, e.g., 0.5°C, as a favorable granularity.

[0075] The parameter N is also combined with the current value X, indicated by reference sign 205, which represents the current value of the parameter N set in the production machine 200 N . The parameter N is further combined with the correction value M, indicated by reference sign 206 N , the correction value M is randomly determined and functions to correct the current value X of the parameter N N . Preferably, the correction value M N is defined with k as a natural number, preferably 1, in the range [X N -kS N , X N +kS N , that is, it is randomly determined between a value that is at least one respective step size S N lower than the current value X of the parameter N N and a value that is at least one respective step size S N higher than the current value X of the parameter N N .

[0076] Figure 3 schematically shows a graph 300 of parameters of a production machine, for example the same production machine 200 as in Figure 2, which is being optimized during operation of an embodiment of the method according to the present disclosure, for example the same method 100 as in Figure 1.

[0077] The graph 300 shows the parameter x 301 of the production machine on its horizontal axis, and shows the function f 302 of the parameter x 301 on its vertical axis, which represents the (partial) influence of the parameter x 301 on the output of the production machine. The graph 300 further shows the current value X 303 of the parameter x 301 that results in an output influence f(X) 304.

[0078] During the operation of an embodiment of the method according to this disclosure, the modified value M 307 is randomly determined within the range from X-kS 306 to X+kS 305, i.e., from a value at least one step size S lower than the current value X of parameter x to a value at least one step size S higher than the current value X of parameter x. The parameter x of the production machine can then be set to this modified value. The resulting output effect f(M) 308 can be evaluated, and a linear gradient L representing an approximation of the derivative of parameter x can be determined based on f(X) 304 and f(M) 308. Based on the linear gradient L, estimations of estimated output effects f(X-kS) 311 and f(X+kS) 310 may be made (indicated by an X). Based on which of these output effects is most likely to result in an overall improved effect on the output of the production machine, parameter x can then be modified by at least its step size S.

[0079] Figure 4 schematically shows, in four graph frames, an example of a loss function used in one embodiment of the method according to this disclosure, for example, the same method as shown in Figure 1.

[0080] The exemplary loss function is applied to a specific example of the stretch blow machine described above, taking into account the material thickness of the base, bottom, middle, and top of the multiple bottles, as shown here.

[0081] Each dot represents a bottle (due to the large number of bottles, adjacent dots in the figure appear filled in, like the thicker parts of a curve). The figure shows that the sigmoid form is used as an example loss function, but other forms may also be used.

[0082] In each frame, horizontal lines are also shown, and the sum of the heights of these horizontal lines represents the value of the loss function for multiple bottles, indicating the value at which the loss function is optimized. As can be seen from the figure, the loss function in this particular example provides an optimization process with relatively high flexibility with respect to the base of the bottle, which allows for better optimization of other parts of the bottle. This is because, at the base of the bottle, the thickness can be reduced (i.e., the dot can shift to the left on the horizontal axis) with only a small change in the result of the loss function.

[0083] Figure 5 schematically shows, in six graph frames, an example of six parameters of a production machine, for example, the same production machine 200 as in Figure 2, during the operation of one aspect of the method according to the present disclosure, for example, the same method as in Figure 1.

[0084] In the diagram, the top three boxes represent passive parameters, i.e., parameters that cannot be modified by the optimization method, while the bottom three boxes represent active parameters, i.e., parameters that can be modified by the optimization method.

[0085] Note that during the first few minutes of the illustrated experiment, the active parameters remained unchanged, appearing as a horizontal line until around 12:36.

[0086] As can be seen from the figure, the optimization method iteratively modifies the active parameter by adjusting it by each step size, searching for the optimal state. In this particular experiment, each modification was made by the step size, but in other examples, some or all of the modifications may be made by multiples of the step size, if it is advantageous.

[0087] Figure 6 schematically shows, in four graph frames, an example of the distribution of plastic material in a bottle during operation of one aspect of the method according to this disclosure, for example, the same method as in Figure 5.

[0088] Again, in the specific example of the stretch blow machine described above, the first frame from the top of the diagram represents the material thickness of the base portion of the bottle, the second frame represents the middle portion of the bottle, the third frame represents the bottom portion of the bottle, and the fourth frame (i.e., the bottom frame) represents the top portion of the bottle.

[0089] Each frame shows the actually measured thickness as curves 601, 604, 606, and 608, respectively, as part of evaluating the output of the production machine, and the respective linear function models of thickness fitted to the available training data for thickness measurement are shown as curves 602, 603, 605, and 607, respectively.

[0090] Even when the actually measured thickness fluctuates very drastically from time to time, it can be seen that by using a time decay function for sample weights, the linear function model over time, through small discrete corrections, comes to approximate the actually measured thickness very tightly.

Claims

1. A method (100) for optimizing at least one parameter (301) of a production machine (110), - A step (101) of randomly determining at least one correction value for correcting the current value of each of the at least one parameter, the correction value being based on the current value (303) of each of the at least one parameter and a predefined step size (S) for each of the at least one parameter. - In the production machine, the step (102) of correcting at least one parameter to its respective corrected value, - A step (103) of evaluating the output (111, 304) of the production machine, which is the output achieved using the at least one modified parameter, - A step (104) of fitting a linear function model to at least one parameter based on the evaluated output, - The step (105) of estimating at least one effect (308) on the output of the production machine using the linear function model, when each of the parameters is modified by at least the respective predefined step size, - A step (106) of determining whether to modify the at least one parameter by at least its respective predefined step size, based on the desired output of the production machine, taking into account the at least one estimated effect, A method comprising: (107) if it is decided to modify the at least one parameter, modifying the at least one parameter by at least its respective predefined step size.

2. The method according to claim 1, wherein the at least one correction value is randomly determined within a range from a value that is at least one step size lower than the current value of the at least one parameter to a value that is at least one step size higher than the current value of the at least one parameter.

3. The method according to claim 1 or 2, comprising repeating the steps of the method over a plurality of iterations.

4. The method according to claim 1 or 2, wherein the linear function model is fitted to the at least one parameter based on at least one previously evaluated output and at least one previous setting of the at least one parameter of the production machine corresponding to the at least one previously evaluated output.

5. The method according to claim 4, wherein the sample importance weights of the at least one previously evaluated output and the at least one previous setting decrease over time according to a time decay function.

6. The method according to claim 1 or 2, wherein the at least one effect is estimated by approximating at least one derivative of the at least one parameter and determining at least one value of the at least one approximated derivative when each of the parameters is modified by at least its respective predefined step size.

7. The method according to claim 6, wherein the at least one parameter is a plurality of parameters, and the at least one approximate derivative is a gradient.

8. The method according to claim 1 or 2, comprising the step of performing a heuristic configured to determine the optimal combination of modifications of the at least one parameter to approach the desired output, taking into account the at least one estimated effect.

9. The method of claim 8, wherein determining whether or not to modify the at least one parameter involves minimizing or maximizing a loss function, the loss function being based on a sigmoid function and / or the sum of absolute or squared differences between the evaluated output and the desired output.

10. The method according to claim 3, wherein each of the predefined step sizes changes at least once over one of the plurality of iterations.

11. The method according to claim 1 or 2, wherein each of the predefined step sizes is determined for the at least one parameter based on the minimally distinguishable discrete granularity of the at least one parameter in the production machine.

12. A computer program comprising instructions configured to perform the method according to claim 1 or 2 when executed on a computer processor.

13. A computer program product comprising a computer-readable medium for storing the computer program described in claim 12.

14. A computer device comprising the computer program product described in claim 13, and configured to perform the method described in claim 1 or 2.

15. A production machine (110, 200) operably coupled to the computer device described in claim 14.

Images