System and method for improving configuration of made-to-order designs

Abstract

The invention relates to a method and a system for determining a set of output configuration values characterizing a dedicated configuration of a complex product, the method comprising: a) receiving (201) a set of input configuration parameters; b) providing (202) at least a part of the input configuration parameters as input to a solver; c) the solver computing (203) at least one output value from the provided input configuration parameters; d) determining (205) a set of output configuration values from at least said output value computed by the solver; characterized in that the solver is configured for solving a first-order logic function encoding an algorithm of a trained deep neural network (hereinafter DNN), wherein the algorithm of the DNN has been trained for modeling a function T of an external configuration tool (hereinafter ECT), the function T of the external configuration tool being required for determining a dedicated configuration of a complex product.

Description

[0001] This invention relates to an ETO (Employment-to-Order) configuration system and method.

[0002] An ETO (Entity Parameter Origin) is a structured set of all parameters that constitute a valid specification for a complex product. The complexity of parameter dependencies distinguishes ETOs from simpler (primarily Boolean) product configurations. In simpler configurations, most rules are easily expressed using propositional logic, while ETO configuration rules, which manage available parameter values, typically involve external tools such as simulators, database lookups, or the execution of arbitrary code. Examples of products manufactured according to ETO configurations range from tubular heat exchangers on gas turbines and military and civilian trucks to conveyor belt systems.

[0003] Today, the standard for classic product configuration environments is to allow users to set configuration options in any order. Specifically, users might start by setting the most important parameters for a particular use case, while the tools used for product configuration automatically determine the valid set of values ​​for the remaining (and less important) parameters. Technically, due to recent advances in satisfiability checks (SAT), the ability to specify parameters in any order or to have irrelevant parameters automatically derived based on some optimality criterion has become feasible.

[0004] However, these techniques do not transfer to ETO configuration: in all state-of-the-art ETO configuration tools, users are guided through the configuration process step-by-step. Specifically, users are neither allowed to skip irrelevant configuration parameters nor to start the configuration process with the most important parameters. Instead, the configuration tool enforces that all parameters be provided in a strictly predefined order, regardless of whether that order matches the user's priorities.

[0005] As an example, consider the ETO configuration of a military truck. For the first user, the truck's weight may be irrelevant, while the strength of the armor is crucial. Therefore, the first user needs to specify the armor strength first, and expects the configuration tool to limit the remaining parameter values ​​(e.g., maximum speed, engine power, space) selectable in the remaining configuration workflow to conform to the specified armor strength. For the second user, weight may be crucial, as the truck is intended for transport by aircraft, for example. Naturally, the second user needs to provide the maximum weight first, as an initial condition, and expects the configuration tool to check that the remaining selection of parameter values ​​does not violate the initial condition regarding the truck's permissible weight.

[0006] While such unordered configurations are available in most state-of-the-art Boolean allocators, they are not supported by any modern ETO configuration system. For example, if r(x1,x2,...,x...)... n ,y1,y2,...,y m) is an ETO rule whose validity is determined by an external tool T (e.g., simulation run), which expects input parameters x1,...,x n And it provides output values ​​y1,...,y m The fact that T hides the relationship between its inputs and outputs makes any rule r that depends on T unsuitable for unordered configurations.

[0007] Assuming the user specifies y k A certain condition C. Any tool that supports unordered configuration must derive the input set {(x1,...,x...} n )∣T(x1,...,x n )=(y1,...,y m )∧C(y k = True}. However, the mapping implemented by T remains opaque, and T generally does not allow backward reasoning. Therefore, it is not easy to derive the input set S of a specific output by executing T.

[0008] Several attempts have been made to address the problems associated with ETO configuration by extending the methods used in Boolean configurators to ETO. However, these techniques are too limited to be practical because they fail entirely if the validation of some rules requires invoking an external simulator that depends on more than two or three discrete parameters. Validating any such rule involves at least one invocation of an external solver (e.g., a finite element method (FEM) solver). Typically, many such invocations are necessary to determine the permissible combinations of remaining configuration options that respect the user's previous choices (this can take several hours). Each rule of the ETO system, executing code or an external tool during validation, is evaluated multiple times each time the user sets parameters. In practice, combination expansion requires enabling external tools thousands of times whenever the user makes a single choice. This is unscalable and impractical due to the significant computational time consumed for each configuration step.

[0009] The purpose of this invention is to provide an ETO configuration system and method that allows users to freely choose the order in which parameters are set in complex ETO configuration tasks, while reducing the time required to implement the configuration process, thus making the ETO configuration process for complex products more efficient compared to the prior art.

[0010] According to the invention, this objective is achieved by an ETO configuration method and system for efficiently determining a set of output parameters characterizing a complex product from a set of input parameters, as required by the independent claim. The dependent claims present further advantages of the invention.

[0011] This invention relates to an ETO configuration method for automatically determining an output configuration value set based on an input configuration parameter set, wherein the output configuration value set is configured to define a dedicated configuration for a complex product that satisfies specific predefined requirements or constraints. The determination of the dedicated configuration involves an external configuration tool (hereinafter referred to as ECT), which is configured to provide ECT output values ​​(or several ECT output values) based on one or more ECT input parameters. The products considered in this invention are typically complex products, i.e., whose configuration requires enabling / invoking ECT. Therefore, a complex product according to this invention refers to any product (e.g., simulator, ODE solver, database, numerical computation, solution of differential equations, etc.) where solving the product configuration problem involves evaluating the dependency on ECT. In the field of determining the configuration of such complex products, the method according to this invention includes:

[0012] - For example, an input configuration parameter set is received via a first interface, wherein the input configuration parameters included in the set are configured to specify features required to determine the specific configuration of the product, preferably all features (in other words, the input configuration parameter set includes some or all parameters required to configure the product, i.e., to determine the specific configuration of the product). In particular, different input configuration parameters of the set may be received temporarily at different times during configuration: for example, a subset S T1 Received at time T1 and including input configuration parameters provided by the user and optionally by the system according to the invention, as well as another subset such as S T2 It can be received at time T2, which occurs after T1 and optionally after running the solver or executing the configuration rules according to the invention, wherein the other subset may include input configuration parameters provided by the user after time T1 and / or from S. T1 The input configuration parameter set includes one or more solver output values ​​obtained from one or more input configuration parameters, and / or one or more parameters obtained by executing configuration rules after time T1. Preferably, the input configuration parameter set includes a first subset and a second subset, wherein the first subset includes one or more input parameters whose values ​​are provided by the user, and the second subset includes one or more input parameters whose values ​​are automatically determined / selected by the system. Typically, the first subset includes input parameters whose values ​​are defined as requirements for a specific configuration, while the second subset includes input parameters whose values ​​are not specified for the specific configuration, and these input parameters can be automatically determined or selected by the system according to the invention, for example, predefined according to material type or product structure, etc.

[0013] - At least a portion of the input configuration parameter set is provided as input to a solver, typically an SMT solver, configured to solve a first-order logic function encoded by an algorithm of a trained deep neural network (DNN), wherein the DNN has been trained to approximate the function required to determine the specific configuration (ECT). The solver is configured to provide a set of output values ​​(including one or more output values) involved in determining the output configuration value from a portion of the input configuration parameter set as output. One or more output values ​​output by the solver are approximations of one or more output values ​​obtained by running the ECT while using a portion of the input configuration parameter set as input. In practice, the DNN according to the invention is configured to approximate the ECT because, when using the same input parameters, the output values ​​obtained by both the ECT and the DNN are sufficiently close such that their difference meets a predefined requirement (e.g., a difference of less than 5%). Preferably, the DNN according to the invention includes a modified linear unit as an activation function, such that the function can be obtained from the input vector (x1,...,x...). n ) to output vector (y1,...,y m A mapping is used to approximate the input-output relationship of ECT, wherein the mapping is a piecewise linear function f: R n →R m The present invention then proposes encoding such piecewise linear functions with first-order logic and using a solver to evaluate their output;

[0014] - The output value set is automatically generated by the solver;

[0015] - The set of output configuration values ​​is determined based on the values ​​output by the solver. For example, the values ​​output by the solver may be transferred (e.g., automatically transferred) to another solver involved in the determination of the specific configuration and used as input to said other solver, or may already be part of the output configuration values ​​included in the set of output configuration values ​​in the sense that they can be used by a manufacturing system for manufacturing complex products.

[0016] - Optionally, the determined set of output configuration values ​​is transmitted (e.g., automatically transmitted) to a manufacturing system for manufacturing (e.g., automatically manufactured) the product, wherein the use of the output configuration values ​​by the manufacturing system results in the manufacture of a complex product according to the specific configuration.

[0017] The present invention also proposes a system for automatically determining an output configuration value set from an input configuration parameter set, wherein the output configuration value set is configured to define a dedicated configuration for a complex product, the dedicated configuration satisfying, for example, predefined requirements or constraints, and the determination of the dedicated configuration for the complex product involves the invocation or activation of at least one ECT. The system includes:

[0018] - The first interface is used to obtain the set of input configuration parameters;

[0019] A computational unit includes a solver configured to automatically compute a set of output values ​​based on at least a portion of an input configuration parameter set. The solver is configured to solve a first-order logic function encoded with an algorithm of a trained deep neural network (DNN), wherein the DNN algorithm has been trained to model a function of an ECT required to determine the specific configuration of the product. This function, for example, enables the simulation of the product configuration based on parameters received as input by the ECT. The function outputs an output value for each simulated configuration of the product, wherein the output value generated by the DNN based on the parameters received as input by the ECT is an approximation of the output value that the ECT would generate when simulating the specific configuration based on the parameters received as input and through the function. The computational unit is further configured to determine output configuration values ​​based on the set of output values ​​computed by the solver.

[0020] - A second interface, which is used to provide the output configuration value determined by the computing unit;

[0021] - Optionally, a manufacturing system, for example connected to the second interface, is configured to prioritize and automatically manage the manufacturing of products based on previously determined output configuration values, so as to produce products according to the dedicated configuration.

[0022] Advantageously, the present invention also proposes a method and system for training a DNN to approximate ECT, the method comprising:

[0023] - Receives a set of input parameters, where each set includes all parameters required by the ECT function to determine the product configuration, i.e., the associated set of one or more ECT output values ​​involved in determining the specific product configuration for a complex product;

[0024] - For each set of input parameters, receive one or more associated ECT output value sets, wherein the associated set of ECT output values ​​has been obtained by running ECT using the set of input parameters associated with it as input to the latter;

[0025] - Train an algorithm for a DNN that approximates an external configuration tool, wherein the training is based on a set of input parameters and a corresponding set of associated ECT output values;

[0026] - Store the trained DNN.

[0027] The aforementioned method can be performed by a training system according to the present invention, wherein the training system comprises:

[0028] - The first training interface, which is configured to obtain the set of input parameters;

[0029] - A second training interface, which is configured to receive the associated set of ECT output values ​​for the execution of the function from ECT for each set of input parameters;

[0030] - A training computation unit configured to train a DNN based on each pair of sets, each pair of sets including a set of input parameters and its associated set of ECT output values, wherein the training computation unit is configured to control the ECT such that the ECT provides a set of ECT output values ​​for each set of input parameters, the set of ECT output values ​​being collected by the training computation unit via the second training interface for training the DNN;

[0031] - A memory configured to store the trained DNN.

[0032] Preferred, but not exclusive, embodiments of the invention will now be described with reference to the accompanying drawings, in which:

[0033] Figure 1 A preferred embodiment of the system according to the present invention is illustrated schematically;

[0034] Figure 2 A flowchart illustrating the method according to the present invention is shown schematically;

[0035] Figure 3 This is an example of a DNN according to the present invention;

[0036] Figure 4 This is an example of a preferred DNN according to the present invention.

[0037] A preferred embodiment of the system 100 according to the present invention is in Figure 1 As shown in the figure. The system 100 enables a user to easily and efficiently determine the specific configuration of a complex product from a set of input configuration parameters, wherein, when an ECT needs to be invoked, the configuration process (i.e., the process for determining the output configuration values ​​that will characterize the specific configuration of the product) is completely unordered for the user: the system does not force the user to define different aspects of the product according to a predetermined strict order (which is usually unintuitive and common in the prior art when an ECT needs to be invoked), but advantageously allows the user to freely choose which aspects of the product he wants to configure first.

[0038] The system 100 according to the invention typically includes a first interface 101, which includes, for example, a user interface that may include a keyboard or a touchscreen. The first interface 101 enables the system 100 to acquire the set of input configuration parameters. For example, a user may use the user interface to provide a subset of the set of input configuration parameters to the system 100. Alternatively or additionally, the first interface 101 may be connected to a database for automatically acquiring another subset of the input configuration parameters. The input configuration parameters may be, for example, the desired length of a product, and / or a specific weight, and / or one or more materials used for the product, and / or material properties, and / or environmental conditions, and / or a specific geometry, etc.

[0039] The input configuration parameters are then automatically provided to the computing unit 103 of system 100. The computing unit 103 may include one or more processors and optionally one or more memories. The processing unit 103 is configured to automatically determine output configuration values ​​characterizing a specific configuration for a complex product based on the received input configuration parameters. The processing unit 103 may be connected to the second interface 102 and / or the memory. For example, the processing unit 103 may be configured to store the output configuration values ​​in the memory and / or to automatically transmit the output configuration values ​​to the control unit of manufacturing system 104 via the second interface 102. The control unit is specifically configured to control the manufacturing of the product based on the received output configuration values, which define, for example, the material and / or dimensions and / or geometry of the complex product, such that it meets predetermined requirements or constraints. The output configuration values ​​preferably include all necessary information to enable the manufacturing system to technically manufacture the product according to the specific configuration.

[0040] The determination of the output configuration value by the computing unit 102 involves a first phase of data collection as an initial step and a second phase including the training of the DNN. This initial step is preferably performed offline, i.e., offline before the configuration process for a complex product begins. In practice, the present invention proposes executing a function T for processing at least some of the input configuration parameters, such that the computing unit 102 can determine the output configuration value before the configuration process begins, which is typically done online once the user requires a new configuration for the complex product.

[0041] According to the invention, ECT is therefore performed offline before any user invocation of the system according to the invention and based on a large set of input parameters, each set including the input parameters required for executing the function T of the ECT, which is required for determining a specific configuration for a complex product. For each set of input parameters provided to the ECT, the training system according to the invention is configured to provide the set of input parameters to the ECT, execute the function T of the ECT required for determining the specific configuration for a complex product, and collect the set of ECT output values ​​associated with the set of input parameters provided to the ECT via a second interface. In other words, during the data collection phase implemented by the training system according to the invention, the ECT computes the associated set of ECT output values ​​for each set of input parameters via its function T, the set of ECT output values ​​being automatically collected by the training system according to the invention and stored in memory.

[0042] This data collection phase is time-consuming and needs to be completed for each ECT required during the configuration process of a complex product. However, it only needs to be completed once, and can be preferably parallelized, for example, by enabling each ECT to be executed in parallel on a cloud platform with millions of different sets of input parameters, without incurring any loss to any extent.

[0043] Once the data collection phase is complete, or concurrently with it, and as long as the ECT output value set is available, the training system according to the present invention begins training the DNN, including the input parameter set (x1,...,x...). n ) and the associated ECT output value set (y1,...,y m Pairwise data points (x1, ..., x2) are used by the training system to teach algorithm A of the DNN. At the end of training, the trained algorithm A is able to approximate the ECT output value of function T: T(x1, ..., x3) n )=(y1,...,y m ).

[0044] Advantageously, since the function T of ECT is unavailable for the system according to the invention (ECT is a very complex external tool relative to the system according to the invention), the trained algorithm can simulate the behavior of function T without having to run the function thousands of times in a trial-and-error manner. Another advantage is that the DNN has an explicit internal structure capable of performing automated reasoning. In particular, the DNN according to the invention includes a modified linear unit as an activation function, which is configured to formally describe how function T reacts with input (x1,...,x...). n Mapped to ECT output (y1,...,y) m An approximation of ). By constructing, such as Figure 3 As shown, the variables x1,...,x nThe input neurons of a 301 DNN and their variables y1,...,y m The output layer 302 and the weights and biases of its hidden layer 303 provide an approximation of the function T obtained by the algorithm for training the DNN.

[0045] Therefore, instead of executing function T thousands of times during ETO configuration to determine (x1,...,x) n ) and (y1,...,y m The system according to the invention uses a DNN to approximate the function T, which is required to determine the specific configuration of the product, based on the dependencies between T and ECT. Compared to ECT, the DNN can potentially execute almost instantaneously, i.e., several orders of magnitude faster than a call to the function T.

[0046] Combination Figure 4 This presents another advantage of the DNN according to the invention, wherein the relationship between layer i-1401 and layer i-402 is presented: the DNN according to the invention implements a mapping by function T in a manner easily accessible by solver techniques. An approximation. In fact, by considering the k-th neuron in the i-th layer, its value... The sum of neurons in layer (il) is determined by the weight matrix W of layer i. i ∈R Ni×Ni-1 It is weighted by its activation function σ. Formally, According to the invention, σ is preferably a modified linear unit, i.e., σ(x) = max(x, 0). Therefore, and advantageously, the value of each neuron in the DNN... It is determined by a piecewise linear combination of its inputs, and therefore by a piecewise linear function that makes the DNN with modified linear units an activation function accessible to the satisfiability solver. In fact, such a function can be directly encoded in first-order logic and fed into the decision process of the solver according to the invention.

[0047] Advantageously, a solver is used to evaluate a first-order logic function that encodes a piecewise linear function of a DNN that approximates the ECT function T, enabling evaluation with respect to the mapping. Reasoning can proceed in any order. Specifically, given the formalization of a piecewise linear function f in first-order logic, the rule r(x1,...,x...) is... n ,y1,...,y m The satisfiability can be automatically evaluated by the satisfiability solver without executing the function T of ECT. Furthermore, the evaluation is undirected, meaning the solver can evaluate with respect to (x1,...,x...). n The constraints of ) propagate to (y1,...,y) mThe constraints, and the solver can solve problems related to (y1,...,y). m The constraints of ) propagate to (x1,...,x) n This constraint allows for completely disordered ETO configuration, providing a significant advantage over existing technologies.

[0048] Figure 2 A flowchart 200 is presented schematically illustrating the steps of a method according to the invention, which can be performed by the previously described system 100. Figure 2 The steps occur after the initial steps of the first phase, which includes data collection, and the second phase, which includes DNN training.

[0049] At step 201, system 100 receives a set of input configuration parameters. For example, a user can provide input configuration parameters required to determine a specific configuration of the product via the first interface 101. Advantageously, system 100 can also receive one or more output configuration values ​​as "inputs," which the product must meet or satisfy once the product configuration is implemented (in other words, system 100 enables the user to provide y i The value, and can be determined according to y i The value of the input configuration parameter is automatically determined (constraints or requirements on the input configuration parameter). Other input configuration parameters can be automatically determined by the system 100, for example, from a database. At different times during the process of determining the specific configuration for a complex product, the system may receive input configuration parameters optionally requested by the user.

[0050] At step 202, the system is configured to provide some or all of the input configuration parameters as input to the solver. As previously explained, the solver is configured to solve for a first-order logic function encoded by a DNN algorithm, wherein the DNN has been trained to approximate the function T of the ECT.

[0051] At step 203, the solver calculates or generates an output value based on the input configuration parameters it receives as input.

[0052] At step 204, and optionally, some or all of the previously computed or generated output values ​​are used as input to another solver configured to solve another first-order logic function encoded by an algorithm of another DNN, wherein the other DNN has been trained to approximate another function T' of an ECT or another function T' of another ECT. Alternatively or additionally, some or all of the previously computed or generated output values ​​are used by the system as input for executing configuration rules. As described in step 203, the other solver computes or generates output values ​​based on the input configuration parameters it receives as input. Similarly, the execution of configuration rules enables the system to determine one or more additional output values ​​for determining the specific configuration of a complex product. When executing another solver and / or another configuration rule involved in determining the specific configuration of a complex product, the output values ​​generated by the solver and / or configuration rule can be further used by the system as input configuration parameters, with or without the additional input configuration parameters.

[0053] At step 205, the system determines the set of output configuration values ​​based on one or more values ​​output by one or more solvers and / or one or more configuration rules. The output configuration values ​​typically define a product-specific configuration based on predefined requirements provided to the system, such as as input configuration parameters. For example, the system may use some or all of the input configuration parameters and values ​​output by one or more solvers to calculate a specific shape or size of the product, or to determine the material to be used, the wall thickness, or the spatial reallocation of weld points, etc.

[0054] At step 206, and optionally, the system according to the invention automatically transmits the output configuration values ​​to a manufacturing system comprising a manufacturing machine configured to manufacture a product according to the output configuration values. For example, all or part of the output configuration values ​​may be used to control the manufacturing machine involved in the technical manufacturing of the product, wherein the use of the output configuration values ​​produces a product characterized by the specific configuration, the output configuration values ​​enabling, for example, specialized control of the movement of manufacturing tools of the manufacturing machine.

[0055] Compared with existing solutions, the present invention presents the following advantages:

[0056] - It allows users to choose the order in which input configuration parameters must be set in complex ETO tasks. Additionally, the training of the DNN by the training system according to the invention is separate from the configuration process handled by the system according to the invention, and therefore training can be performed well before any configuration process begins, and can be easily done in a massively parallel manner. Compared to all state-of-the-art methods that execute external code / simulators during the configuration process, using a DNN speeds up the configuration process by several orders of magnitude: no external tools need to be executed during configuration;

[0057] - It proposes using a DNN with modified linear units to obtain a good approximation of ECT through training, so that the resulting DNN falls into the category of piecewise linear functions, which can be efficiently analyzed using decision processes (such as SMT solvers, MILP solvers, CPLEX, or delta-decision processes such as dReal).

Claims

1. A method for designing configurations to order for determining a set of output configuration values ​​representing a specific configuration of a complex product, the method comprising: a) Receive (201) an out-of-order set of input configuration parameters; b) Provide at least a portion of the (202) input configuration parameters as input to the solver; c) The solver calculates at least one output value (203) based on the provided input configuration parameters; d) Determine (205) the set of output configuration values ​​based on at least the output values ​​calculated by the solver; The feature is that the solver is configured to solve a first-order logic function that encodes an algorithm for a trained deep neural network (DNN), wherein the DNN algorithm has been trained to model a function T of an external configuration tool (ECT), the function T of which is required to determine the specific configuration of the complex product. The first-order logic function, which encodes the algorithm of the trained DNN, enables the solver to solve the first-order logic function using at least a portion of the unordered set of input configuration parameters.

2. The method according to claim 1, wherein, The at least one output value is used as input (204) in another solver configured to solve another first-order logic function that encodes an algorithm for another DNN, wherein the other DNN has been trained to approximate another function T' or another function T" of ECT.

3. The method according to claim 1 or 2, further comprising automatically transmitting (206) an output configuration value to a manufacturing system configured to control a manufacturing process based on the output configuration value.

4. The method according to claim 1 or 2, wherein, The DNN includes a modified linear unit as an activation function.

5. A method for training a DNN to approximate a function of ECT in a make-to-order design configuration method that determines a set of output configuration values ​​characterizing a specific configuration of a complex product, the method comprising: - Receive a set of input parameters, wherein each set includes all parameters required by the first-order logic function of the ECT to compute one or more associated sets of ECT output values ​​involved in determining the specific product configuration for a complex product; - For each set of input parameters, receive the associated set of one or more ECT output values, wherein when the set of input parameters associated with one or more ECT output values ​​is used as the input to the ECT, the one or more ECT output values ​​have been obtained by running the first-order logic function of the ECT; - Train an algorithm for a DNN that approximates the ECT, wherein the training is based on pairs of sets including each set in the input parameter set and its associated set of output values; - Store the trained DNN; The trained DNN has an algorithm encoded by the first-order logic function, such that the first-order logic function can be solved by a solver that takes at least a portion of an unordered set of input configuration parameters as input.

6. A make-to-order design configuration system for determining a set of output configuration values ​​characterizing a specific configuration of a complex product, the system comprising: - First interface (101), which is used to obtain an unordered set of input configuration parameters; - Computation unit (103), which includes a solver that is set with at least a portion of an unordered set of input configuration parameters as input; - A second interface (102) is used to provide the output configuration value generated by the computing unit (103); The system is characterized in that the solver is configured to solve a first-order logic function encoded by an algorithm of a trained DNN, wherein the DNN algorithm has been trained to model a function T of an ECT, the function T of which is required to determine the specific configuration of the complex product. The first-order logic function, which encodes the algorithm of the trained DNN, enables the solver to solve the first-order logic function using at least a portion of the unordered set of input configuration parameters.

7. The system according to claim 6, further comprising: - A manufacturing system configured to automatically control the manufacturing process based on the output configuration values.

8. A training system configured to train a DNN to approximate a function of ECT in a make-to-order design configuration method that determines a set of output configuration values ​​characterizing a specific configuration of a complex product, the training system comprising: - The first training interface, which is configured to obtain the set of input parameters; - A second training interface, configured to receive, for each set of input parameters, an associated set of ECT output values ​​generated by executing the function of the ECT; - A training computation unit configured to train a DNN based on each pair of sets, each pair of sets including a set of input parameters and its associated ECT output value set, wherein the training computation unit is configured to control the ECT such that the ECT provides the ECT output value set for each set of input parameters, the ECT output value set being collected by the training computation unit via the second training interface for training the DNN; - A memory configured to store the trained DNN; The trained DNN has an algorithm encoded by a first-order logic function, such that the first-order logic function can be solved by a solver that takes at least a portion of an unordered set of input configuration parameters as input.

Images