Method for controlling and / or monitoring one or more asset transfers

Abstract

The invention refers to a controlling and / or monitoring of an asset transfer. Time series monitoring data related to the asset is provided. A trained data-driven model is provided. The data-driven model partially includes discrete weights, wherein the model has been trained i) by providing a training data set including a) control and / or monitoring data of asset transfers and b) time series monitoring data related to the asset, ii) by initializing the model on a classical computer based on the training data set and iii) by providing the initialized model for determining the discrete weights based on the training data set to a quantum computer. The trained model determines control and / or monitoring data for controlling and / or monitoring an asset transfer based on the time series monitoring data. Control and / or monitoring data are generated based on the trained data-driven model and the time series monitoring data related to the asset.

Description

[0001] Method for controlling and / or monitoring one or more asset transfers

[0002] FIELD OF THE INVENTION

[0003] The invention refers to a computer-implemented method, an apparatus and a computer program product for controlling and / or monitoring one or more asset transfers. Moreover, the invention refers to a training method, a training apparatus and computer program prod- uct for training a data-driven model usable in the method apparatus and computer program product for controlling and / or monitoring one or more asset transfers.

[0004] BACKGROUND OF THE INVENTION

[0005] In the modern producing industries, for instance, chemical industries, asset transfers play a crucial role in the development and improvement of production processes. For example, carbon credits that are based on a measured carbon emission are crucial for the reduction of carbon emissions in production processes. Thus, improving a controlling and / or monitoring of asset transfers can play an important part in optimizing a production, for instance, of a chemical product, with respect to the real-world aspect represented by the asset.

[0006] SUMMARY OF THE INVENTION It is an object of the present invention to provide an apparatus, a method and a computer program product that allow to control and / or monitor asset transfers in real-time. Moreover, it is an object of the present invention to provide a training apparatus, training method and computer program product allowing to train a data-driven model that can control and / or monitor an asset transfer in real-time. BASF SE 230213

[0007] Today, the production industries, for instance, the chemical industries, act, in particular, produce, in a complex network of producers, vendors, service providers and consumers that participate in the value chain of a product. Improving a production of a product in the context of this complex network of different participants in the product value chain is difficult as in many cases the producer of the product only has influence on a small part of this complex network. For example, a product producer might be able to optimize certain steps in the production of the product to reduce a CO2 emission, however, has no direct influence on the CO2 emission of providers of pre-products, logistic providers or of the consumers of the produced product. In this complex environment asset transfers are a possibility to ensure that certain aspects are optimized with time throughout the whole of the product value chain without directly influencing or regulating the production. For example, CO2 certificates are generated and transferred between participants of a product value chain and indirectly lead all participants of the product value chain and also CO2 transfers to optimize their production with respect to the CO2 emission. In another example an asset value can be indicative of an aspect with respect to which a production should be optimized. For instance, a value of an energy asset tends to be low when the energy is produced by green energy sources producing no or less CO2. Thus transferring energy assets during such times allows to optimize a production of a product to save CO2. Instead of a CO2 emission reduction also other goal can be optimized by monitoring and / or controlling asset transfers, for example, packaging, biodegradation, recycling, resource consumption, renewable energy and energy consumption, wherein each of this goals provide respective asset transfers that influence the respective goal. Accordingly, controlling and / or monitoring of asset transfers is important for the producing industries and for optimizing their respective product value chain with respect to one or more aspects represented by the asset.

[0008] However, controlling and / or monitoring asset transfers provides its own technical challenges, in particular, with respect to the timing. Moreover, the respective asset transfer environment can be complex providing a plurality of influences and parameters such that developing rules for the automatic controlling and / or monitoring, in particular, in view of the to be utilized computational resources can be challenging.

[0009] Since the below described method includes utilizing a trained data-driven model that is configured to be trainable by a quantum computer in a hybrid quantum computational approach, the quantum computation as part of the training allows the inclusion of more of the complexity of the environment of the asset transfer into the trained data-driven model in a computationally effective way. Moreover, a such trained data-driven model allows for controlling and / or monitoring of asset transfers in real-time. Particularly advantageous are the discrete weights utilized by the data-driven model and trained by the quantum computing BASF SE 230213 unit that allow for utilizing an effective utilization of quantum computational resources, in particular storage and calculation requirements that facilitate the real-time aspect and at the same time allow the deployment of the controlling and / or monitoring even to computational devices with limited or restricted computational resources like mobile devices.

[0010] In a first aspect a computer-implemented method for controlling and / or monitoring one or more asset transfer(s), wherein the method comprises I) provide time series monitoring data related to the asset, II) provide trained data-driven model, wherein the data-driven model at least partially includes discrete weights, wherein the data-driven model has been trained i) by providing a training data set including a) control and / or monitoring data of asset transfers and b) time series monitoring data related to the asset, ii) by initializing the model on a classical computer based on the training data set and iii) by providing the initialized model for determining the discrete weights based on the training data set to a quantum computing unit, wherein the trained model determines a model output indicative of control and / or monitoring data for controlling and / or monitoring an asset transfer based on the time series monitoring data, III) generate control and / or monitoring data based on the trained data-driven model and the time series monitoring data related to the asset, and IV) provide control and / or monitoring data for controlling and / or monitoring a transfer of the asset.

[0011] The method is a computer-implemented method that can be carried out, for instance, by carrying out a computer program product on any form of computational hardware. In particular, the computer-implemented method can be realized in form of any combination of software and / or hardware that provides the functions defined by the method. For example, the hardware can refer to any known dedicated or general classical computer hardware. For example, the computer-implemented method can be carried out on any known computational device, like a personal computer. However, the computer-implemented method can also be realized in a cloud environment, computational network, etc., such that at least parts of the method, for instance, one or more functions defined by the method, are realized in a network solution and thus spread over a plurality of computational devices.

[0012] The asset is a tangible or intangible resource that can be associated with the production of a product and / or more generally with the value chain of a respective product. Furthermore, the asset can include a digital representation of the asset, for instance, associated with a production and / or product performance of a product. For example, a digital representation of an asset related to a product performance with respect to an environmental impact of a product can be a carbon credit or certificate that is based on a measured carbon emission. For instance, a carbon credit can be determined based on a measured carbon emission reduction of a production process compared to a reference carbon emission and is provided BASF SE 230213 in form of a digital representation of the respective carbon emission reduction. Further, the digital representation of an asset can refer to a digital representation of a recycled content of a material or product produced at least partly from recycled material. A similar example refers to a bio-based content of a material or product at least partly produced from biological materials. In these cases, respective certificates can represent the respective content, and, for instance, certify the percentage of the product related to the content. Moreover, the digital representation of the asset can refer to a certificate over a biodegradability of a material or product. The digital representation of an asset can be any representation that can be processed by a computer system and represent the respective asset. For example, the digital representation can be a data package associated with a unique identifier and including respective asset data. The digital representation of the asset can then be stored in a secure data storage structure such as a wallet. Such a digital wallet can be an electronic device, online service or software program that allows participating in a respective asset transfer, for instance, by providing a platform for the electronic management of the asset transactions and can further provide a respective authentication of the participant of a respective transaction. The transfer of an asset can be recorded as transaction on a distributed ledger, for instance, a Hyperledger, blockchain such as Ethereum, Solana, Bitcoin etc. Utilizing a distributed ledger for recording respective transactions and transfers of assets allows for a decentralization and improved security for all participants of the asset transfer. An asset transfer refers in this context to a transfer of an ownership of an asset, for instance, represented by the ownership over the digital representation, from one participant of the transfer to another. For example, the transfer of the ownership can be recorded as asset transfer in the digital representation of the asset, for instance, as part of a blockchain representing the asset.

[0013] In a step the method comprises providing time series monitoring data related to the asset. Time series monitoring data comprises data points of respective monitoring data related to the asset that are associated with a time stamp indicating a point in time in which the monitoring data was recorded, for instance, measured. The time series monitoring data can include any information that is related to the asset. For example, the time series monitoring data can refer to measurement data related to one or more quantities related to the asset that can be measured. For instance, with respect to carbon dioxide emissions the time series monitoring data can relate to a measured reduction or more generally to a measurement of the carbon dioxide emission. In a particularly advantageous example, the time series monitoring data related to the asset is associated with the transfer value of the asset transfer. For example, the transfer value can be associated with the computing cost of an asset transfer, like Ethereum gas fee, or can be associated with an equivalent value for BASF SE 230213 performance assets like carbon credits or credits referring to utilizing recycled materials in production.

[0014] In a next step, a trained data-driven model is provided. The term “data-driven” is used herein to emphasize that the model is mainly based on respective data input and not, for instance, on intuition, personal experience or knowledge. For example, the trained data- driven model refers to a machine learning based model that is based on known machine learning algorithms, like neural networks, regression models, classification algorithms, etc. In particular, the data-driven model includes a set of discrete weights. The weights of a trained data-driven model represent the training of the model and thus the trained functional relation between the input and output data to the trained data-driven model. Thus, the weights can be regarded as representing the information extracted from the training data utilized during the training process of the trained data-driven model. The data-driven model is trained based on a provided training data set including a) control and / or monitoring data of asset transfers and b) time series monitoring data related to the asset. The time series monitoring data related to the asset can be regarded as features and the control and / or monitoring data of asset transfers as labels to the features such that the training data set forms a labelled data set for training the data-driven model. The control and / or monitoring data can be “real” control and / or monitoring data that has been used for the respective tasked based on the time series monitoring data, for instance, by an operator, or it can be synthetic or theoretical control and / or monitoring data that is derived from the time series monitoring data derived to optimize a respective goal of the controlling or monitoring based on the full knowledge of the time series monitoring data. Moreover the labels can also refer to quantities that can be derived from the control and / or monitoring data. For example, the labels can refer to a CO2 reduction archived by a respective production controlling action or an asset value difference derived from a respective asset trading controlling action. Further the labels can also refer to parameters from which respective controlling and / or monitoring action can be derived. For example, such a parameter can refer to a development of a quantity of the time series monitoring data a predetermined time after the data used as feature, e.g. whether a quantity increases or decreases after the respective feature data time. In such cases the controlling and / or monitoring data is derivable from the parameters based on predetermined or also learned rules. The data of the training data set can refer to historical data, for instance, to past-time series monitoring data that has been recorded together with the control and / or monitoring data of asset transfers associated with the time series monitoring data. However, the training data set can also include synthesized data generated based on simulations, theoretical considerations or known functionalities in association with the controlling and / or monitoring of asset transfers and associated time series monitoring data. BASF SE 230213

[0015] Further, the training of the data-driven model includes initializing the data-driven model on a classical computer based on the training data set. For example, the initializing can refer to a pre-training of the data-driven model based on the training data set or a part of the training data set. Moreover, the initialization can refer to generating an optimization function based on the training data set that can be optimized on the quantum computing unit for training the discrete weights of the data-driven model. Further, the training includes providing the initialized model for determining the discrete weights based on the training data set to a quantum computing unit. A quantum computing unit refers to a unit that performed the quantum computation utilizing the physics of quantum mechanics for operating the computation. Quantum computing units are based on quantum elements adhering to the physics of quantum mechanics, such as superconductors, ions, atoms, quantum dots, photons, particle spins, bosons etc. During a quantum computation, these quantum elements can be manipulated in a controlled manner to perform operations that follow the rules of quantum mechanics. Thus, a quantum computing unit is fundamentally different from a classical computing unit in that it provides operation possibilities going beyond the operation possibilities that can be followed by an electronic device like a classical computer. For computing a problem, for instance, for training the data-driven model utilizing the initialized model on a quantum computing unit, the respective problem, for instance, the training, has to be formulated in quantum mechanical terms to be mapped on the quantum elements of the respective quantum computing unit. Generally, translations of a wide variety of problems to quantum mechanical problems are well-known and can be utilized to translate the training of the data-driven model based on the initialized model and the training data set to a quantum computing unit. Such a translation can be part of the providing of the initialized model. However, a translation can also be performed by a control unit of the quantum computing unit that controls and monitors the quantum computation on the quantum computing unit, for instance, that manipulates the respective quantum elements utilizing a respective quantum algorithm. In such a case, the initialized model together with the training data set and comprising the problem to be solved is provided to the quantum computing unit and the control unit of the quantum computing unit, being a classical computing unit, translates and prepares the problem such that the training is performed on the quantum computing unit.

[0016] It has been found by the inventors that a particularly advantageous data-driven model is a boosting classifier comprising a plurality of weak classifiers. In this example, the discrete weights determine respective contributions of classifications of the weak classifiers to the classification of the boosting classifier. A weak classifier can be machine learning based model for binary classification that performs slightly, for example, with a predetermined amount below a predetermined threshold, better than random guessing. This means that BASF SE 230213 the weak classifier will provide an output that is known to have some skill, e.g. making the capabilities of the model weak, although not so weak that the model has no skill, e.g. performs worse than random. A weak classifier can also be a non-binary classifier that performs slightly better than a naive determination method. The initialization of the model can then include training the weak classifiers based on the training data set on a classical computing unit and generating based on the results of the trained weak classifiers an optimization function that determines the discrete weights that determine the contribution of the results of the weak classifiers. This respective optimization function can then be provided as initialized model to the quantum computer and can then be solved, for instance, optimized, based on the training data set to determine the discrete weights for the results of the weak classifiers. The trained data-driven model then refers to the weak classifiers, wherein the results of the weak classifiers are combined utilizing the trained discrete weights to generate the final results, e.g. classification, of the boosting classifier. In an example, the training of the data-driven model can be performed further based on an application target objective. In this case the training not only utilizes the training data sets such that the data-driven model determines the model output that matches the labels as well as possible, but also such that the model output matches the application target objective. The application target objective can be any objective that should be met when applying the data-driven model. For example, in case of assets referring to pre-product used for producing a product, the application target objective can be to transfer assets for the production such that the CO2 consumption of the pre-products is as low as possible. Thus the additional target objective can be utilized to optimize a product production with respect to predetermined goals utilizing asset transfers.

[0017] The trained data-driven model when applied to the time series monitoring data as input provides as output data indicative of control and / or monitoring data for controlling and / or monitoring a respective asset transfer based on the time series monitoring data. For example, the output of the data-driven model can refer a transfer value of the asset for a predetermined time for monitoring the asset transfer or can refer to an indicator of whether or not to transfer an asset for controlling the asset transfer. The data can refer to an indicator that is predefined to indicate certain monitoring and / or control action. The definition can be provided by convention or based on the most suitable output for the data-driven model and is also utilized for the monitoring and / or controlling data in the training data set. For example, it can be defined that an output of “1 ” indicates to transfer an asset to a wallet of the user, and output of “-1 ” indicates to transfer an asset from the wallet of a user to another wallet and an output of “0” can indicate to not transfer an asset. However, also other conventions or definitions and other output data can be utilized. BASF SE 230213

[0018] In a further step, control and / or monitoring data are generated based on the trained data- driven model and the time series monitoring data related to the asset. For example, the trained data-driven model is applied to the time series monitoring data by providing the time series monitoring data as input to the trained data-driven model. The trained data-driven model then provides as output an indicator based on which the control and / or monitoring data can be generated. For example, the indicator can be a classification like a discrete value that indicates to control a transfer of an asset. Additionally or alternatively, the indicator can be indicative of a transfer value, for instance, can be indicative of a rising or falling of a value of an asset, at a predetermined time period for monitoring the asset transfer. The respective control and / or monitoring data can then be generated with respect to the system utilized for the transfer of monitoring of the asset. For example, the control and / or monitoring data can comprise respective computer operations and instructions that case a classical computer to perform the control and / or monitoring action indicated by the output of the data-driven model.

[0019] In a further step, respective control and / or monitoring data can then be provided for controlling and / or monitoring the transfer of the asset. For example, the control and / or monitoring data can be provided to a user for verification or information together with respective suggestions on a further procedure, wherein the user can then in a machine / user guided interaction control and / or monitor the transfer of the asset. However, the control and / or monitoring data can also be configured to allow for an automatic controlling and / or monitoring of the asset transfer without further user intervention. For instance, the control data can be configured to initiate a transfer of the asset if indicated by the results of the trained data-driven model. Automatic monitoring can, for instance, include flagging predetermined events in the time series monitoring data based on the results of the trained data-driven model and, for instance, notifying a user of these events.

[0020] In an embodiment, the training of the data-driven model on the quantum computer comprises removing correlations in the initialized model above a predetermined threshold. The removed correlations can refer to any correlations of features or other parts of the model. For example, if the model refers to a boosting classifier comprising a plurality of weak classifiers, the correlations can be determined with respect to the contributions of classifications of the weak classifiers to the classification of the boosting classifier and contributions correlating above a predetermined threshold can be removed. For example, a correlation can be determined between respective two results of weak classifiers during a classification and compared to a threshold, wherein if the correlation lies above the predetermined threshold, i.e. if the results of the classifications show a predetermined similarity, one of the respective weak classifiers can be removed, for instance, by determining its weight and BASF SE 230213 thus its contribution as zero or by completely removing the weak classifier. Utilizing a quantum computer for this task during the training of the data-driven model, in particular, during the determination of the discrete weights of the weak classifiers, allows to take into account a larger set of weak classifiers and thus characteristics, information and features of the time series monitoring data than would be possible with any classical computer. At the same time, the removing of the correlations allows for a very computer resource efficient final trained data-driven model that can then be performed by a classical computer.

[0021] In an embodiment, the method further comprises aggregating the received time series monitoring data for a predetermined time interval into a feature set representing the time series monitoring data for that predetermine time interval and wherein the data-driven model is trained to generate the model output based on the feature set for the predetermined time interval as input. For example, the received time series monitoring data can be split into time intervals with a predetermined length and the feature set can be aggregated from the time series monitoring data for each of the time intervals. However, the features can also be aggregated for selected or predetermined time intervals of the time series monitoring data. The feature set represents the time series monitoring data of a respective time interval by providing respective features that characterize the time series monitoring data in the respective time interval. For example, the feature set of a predetermined time interval comprises as feature at least one of: a simple moving average, a relative strength index, a rate of change, and a moving average converge divergence. However, also other kinds of features can be used advantageously and the utilized features can even depend on the respective application. The features utilized in the feature set can be predetermined by a user, for instance, based on experience, theoretical considerations, etc. However, the features utilized in a feature set can also be determined as part of the training of the data- driven model utilizing respectively known feature selection algorithms. The data-driven model is then trained to generate the respective model output indicative of control and / or monitoring data based on the respective feature set for a respective time interval. Accordingly, control and / or monitoring data can be associated with the respective time interval of the time series monitoring data. For ensuring a real-time controlling and monitoring, it is advantageous to predetermine the time intervals as short as possible. For example, the time intervals can be predetermined based on the timely density of the time series monitoring data such that a time interval is defined by a predetermined amount of data points within the time interval. This ensures that a reasonable amount of information is utilized to determine the respective features of the feature set. For example, if the time series monitoring data comprises monitoring data recorded per second, a reasonable predetermined time interval could refer to one minute, whereas, if the monitoring data is recorded every minute, a reasonable time interval could refer to 30 minutes or an hour. BASF SE 230213

[0022] In an embodiment, the data-driven model comprises an asset value part comprising at least partly the discreet weights correlated to providing a binary output associated with an asset value, and a control data part for generating the model output based on the binary output of the asset value part. In this embodiment, the data-driven model comprises two parts, wherein the output of the asset value part is at least partly input to the control data part. The asset value part comprises the discrete weights that are in this embodiment trained and thus correlated with a binary output associated with an asset value. For example, the binary output can indicate whether the asset value is greater or lower than a predetermined threshold or whether the asset value is rising or falling with respect to a previous asset value. The respective binary output is then provided as input to the control data part generating the model output indicative of the controlling and / or monitoring data. The control data part provides a functional correlation between the binary output and the model output. For example the control part can correlate the binary output indicating the asset value with a learned control and / or monitoring action that is indicated by the model output. This functional correlation can be trained, for instance, during the training of the data-driven model based on the training data. However, this functional relation can also be predetermined, for instance, based on respective rules derived from experience of a user or already known functional relations.

[0023] In an embodiment, the method further comprises generating a plurality of data-driven models, wherein each data-driven model comprises a different set of model parameters defining the model, wherein the method further comprises selecting from the plurality of machine learning based models a machine learning based model based on a validation of the data- driven models. The data-driven models of the plurality of data-driven models can be trained either sequentially or in parallel and after training can be stored, for instance, on a respective data-driven model storage. The model parameters defining the model can refer to any parameter of the model that is fixed during the training process and thus defines the structural integrity of a respective data-driven model. For example, the model parameters comprise at least one of: a number of discreet weights, a kind of model architecture, and a regularization parameter. For example, the different model parameters for each respective data-driven model can be randomly chosen or can be selected based on predetermined rules within respective predetermined limits. For each trained data-driven model of the plurality of data-driven models the validation can be performed by determining a performance metric indicative of the performance of the respective data-driven model in the task of determining the model output based on the time series monitoring data. Moreover, also more than one metric, i.e. more than one performing aspect of the respective data-driven models can be determined. The evaluation is then performed based on the one or more determined metrics based on predetermined rules, for instance, based on a predetermined priority of BASF SE 230213 the metrics, and the best performing data-driven model can be selected based on the metrics. The respective metric used for selecting an optimal model can depend on the specific application. Moreover, also additional predetermined selection criteria can be utilized. For instance, models can be selected that use less controlling and / or monitoring actions. Preferably, a selection cost function is utilized for selecting the optimal data-driven model wherein different metrics and / or selection criteria can be provided with predetermined weights for determining the impact of the respective metric and / or selection criterium on the selection. For example, a performance metric can be provided with a high weight in the selection cost function than a criterium referring to the number of controlling actions performed. This then leads to the performance metric being prioritized over a utilized number of controlling actions such that the utilized number of controlling actions only is relevant for the selection in cases in that the performance metric is similar.

[0024] In a further aspect of the invention, a computer-implemented method for generating control data for controlling a quantum computer to generate a data-driven model, wherein the data- driven model is configured to generate control and / or monitoring data suitable to control and / or monitor an asset transfer based on time series monitoring data related to the asset, wherein the method comprises I) provide a training data set comprising a) control and / or monitoring data of asset transfers and b) time series monitoring data related to the asset, II) initialize a data-driven model at least partially including discrete weights, wherein the model is initialized based on the training data set, and III) generate control data to control a quantum computer for training the data-driven model by determining the discreet weights of the data-driven model based on the initialized model and the training data set, wherein the trained model determines a model output indicative of control and / or monitoring data for controlling and / or monitoring an asset transfer based on the time series monitoring data. In an embodiment, the method comprises controlling the quantum computer based on the control data. Further the method can comprise storing the trained data-driven model on a model storage.

[0025] In a further aspect of the invention, an apparatus for controlling and / or monitoring one or more asset transfer(s), wherein the apparatus comprises one or more processors configured to I) provide time series monitoring data related to the asset, II) provide trained data- driven model, wherein the data-driven model at least partially includes discrete weights, wherein the data-driven model has been trained i) by providing a training data set including a) control and / or monitoring data of asset transfers and b) time series monitoring data related to the asset, ii) by initializing the model on a classical computer based on the training data set and iii) by providing the initialized model for determining the discrete weights based on the training data set to a quantum computing unit, wherein the trained model determines BASF SE 230213 a model output indicative of control and / or monitoring data for controlling and / or monitoring an asset transfer based on the time series monitoring data, III) generate control and / or monitoring data based on the trained data-driven model and the time series monitoring data related to the asset, and IV) provide control and / or monitoring data for controlling and / or monitoring a transfer of the asset.

[0026] In a further aspect of the invention, an apparatus for generating control data for controlling a quantum computer to generate a data-driven model, wherein the data-driven model is configured to generate control and / or monitoring data suitable to control and / or monitor an asset transfer based on time series monitoring data related to the asset, wherein the apparatus comprises one or more processors configured to I) provide a training data set comprising a) control and / or monitoring data of asset transfers and b) time series monitoring data related to the asset, II) initialize a data-driven model at least partially including discrete weights, wherein the model is initialized based on the training data set, and III) generate control data to control a quantum computer for training the data-driven model by determining the discreet weights of the data-driven model based on the initialized model and the training data set, wherein the trained model determines a model output indicative of control and / or monitoring data for controlling and / or monitoring an asset transfer based on the time series monitoring data.

[0027] In a further aspect of the invention, a system for controlling and / or monitoring one or more asset transfer(s), wherein the system comprises I) a quantum computer unit, II) an apparatus as described above for generating control data for controlling the quantum computer to train a data-driven model, and III) an apparatus as described above utilizing the data- driven model for providing control and / or monitoring data for controlling and / or monitoring a transfer of the asset.

[0028] In a further aspect of the invention, a computer program product for controlling and / or monitoring one or more asset transfer(s) comprising program code means for causing an apparatus as described above to carry out a method as described above.

[0029] In a further aspect of the invention, a computer program product for generating control data for controlling a quantum computer to generate a data-driven model comprising program code means for causing an apparatus as described above to carry out a method as described above. | BASF SE | 230213 | 230213WQ01

[0030] It shall be understood that the apparatuses as described above, the methods as described above, the computer program products as described above and the system as described above have similar and / or identical preferred embodiments, in particular, as defined in the dependent claims.

[0031] It shall be understood that a preferred embodiment of the present invention can also be any combination of the dependent claims or above embodiments with the respective independent claim.

[0032] These and other aspects of the invention will be apparent from and elucidated with reference to the embodiments described hereinafter.

[0033] BRIEF DESCRIPTION OF THE DRAWINGS

[0034] In the following drawings:

[0035] Fig. 1 illustrates a state representation of a qubit as used in a quantum computing device,

[0036] Fig. 2 illustrates a schematic example of a quantum computing device with qubits as calculation unit,

[0037] Fig. 3 illustrates a schematic example method for generating a control signal to perform operations on a quantum computing device and for processing measurement signals from the quantum computing device,

[0038] Fig. 4 illustrates a schematic example of a hybrid system including a classical and a quantum computing device,

[0039] Fig. 5 illustrates a schematic example of a quantum computing device based on superconductors,

[0040] Fig. 6 illustrates a schematic example of a quantum computing device based on trapped ions,

[0041] Fig. 7 shows schematically an example of the hardware principles of a quantum annealer, | BASF SE | 230213 | 230213WQ01 |

[0042] Fig. 8 shows schematically and exemplarily an architecture of an asset transaction system controlled and / or monitored by an apparatus for controlling and / or monitoring an asset transaction,

[0043] Fig. 9 shows schematically an example workflow for monitoring and controlling an asset transfer using a trained data-driven model,

[0044] Fig. 10 shows schematically an example of a method for controlling and / or monitoring one or more asset transfer(s) using a trained data-driven model and for training the trained data- driven model utilizing a quantum computer,

[0045] Fig. 11 shows schematically an example of a classical-quantum hybrid training algorithm for an exemplary trained data-driven model being a boosting algorithm,

[0046] Fig. 12 shows schematically an example of a dataflow in a method for controlling and / or monitoring one or more asset transfer(s) using a trained data-driven model and for training the trained data-driven model utilizing a quantum computer,

[0047] Fig. 13 shows schematically an example of details of a method for controlling and / or monitoring one or more asset transfer(s) using a trained data-driven model and for training the trained data-driven model utilizing a quantum computer,

[0048] Fig. 14 shows schematically an example of details of a method for controlling and / or monitoring one or more asset transfer(s) using a trained data-driven model and for training the trained data-driven model utilizing a quantum computer, and

[0049] Fig. 15 shows schematically an example of details of a method for controlling and / or monitoring one or more asset transfer(s) using a trained data-driven model and for training the trained data-driven model utilizing a quantum computer.

[0050] DETAILD DESCRIPTION OF DRAWINGS

[0051] In the following first a short introduction into the general basic principles of quantum computers and the performance of calculations of quantum computers will be provided. Further, general principles can also be found in “Quantum Computation and Quantum Information: 10thAnniversary Edition”, M. A. Nielsen and I. L. Chuang (2010). BASF SE 230213

[0052] Classical computing devices use processors which are based on transistors. The state of each transistor has two controllable states 1 or 0 representing a digital binary or a bit. To perform operations on a classical computing device a human readable program code is translated via a compiler into machine-readable instructions. Machine-readable instructions are control signals, e.g. voltage settings, for each transistor. Representations of the machine-readable instructions may include binary or hexadecimal representations. Based on such machine-readable instructions, the operations are performed on the processor of a classical computing device.

[0053] Quantum computation is a relatively new computation method that uses quantum effects, such as superposition and entanglement, to perform certain computations more efficiently than classical digital computers. In contrast to digital computers, which represent information in the form of bits (e.g., “1 ” or “0”), as described above, quantum computing devices, i.e. quantum computers, use qubits, i.e. quantum bits, to represent information. Quantum computing devices are based on quantum elements adhering to the physics of quantum mechanics, such as superconductors, ions, atoms, quantum dots, photons, particle spins, bosons or the like. These quantum elements may be manipulated in a controlled manner to perform operations.

[0054] Although qubits and their manipulation may be described in terms of their mathematical properties, each such qubit may be implemented in a physical quantum element in any of a variety of different ways. Examples of such quantum elements include superconducting materials, trapped ions, photons, optical cavities, individual electrons trapped within quantum dots, point defects in solids (e.g., phosphorus donors in silicon or nitrogen-vacancy centers in diamond), molecules (e.g., alanine, vanadium complexes), or any medium that exhibits qubit behavior comprising quantum states and transitions there between that can be controllably induced or detected.

[0055] Generally, for any given physical quantum element that implements a qubit, any of a variety of properties of that physical unit may be chosen to implement the qubit. For example, if electrons are chosen to implement qubits, then the x, y or z component of an electron spin degree of freedom can be chosen as the property of such electrons to represent the states of such qubits. For any particular degree of freedom, the physical quantum elements can be controllably put in a state of superposition or entanglement and measurements can then be taken in the chosen degree of freedom to obtain readouts of qubit values.

[0056] In contrast to transistors of classical computing devices each quantum element of quantum computing devices can not only take the basis states |1) or |0) but also any superposition BASF SE 230213 of such basis states, such as state |X). The state of each quantum element is represented by a state of a quantum bit, i.e. qubit, as illustrated in the two-dimensional simplification of Fig. 1. To represent such states Dirac notation is commonly used in quantum mechanics. In Dirac notation a state in a n dimensional, complex vector space, such as a Hilbert space, is represented in braket notation, for example |X). According to conventional terminology, the superposition of “0” and “1 ” states in a quantum computing device can be represented as cr|O> + / ?| 1> . The states “0” and “1 ” or bits of the classical computing device are similar to the basis states |0) and |1) or quantum bits of the quantum computing device, respectively. The value |<x|2represents the probability that the qubit will be measured in the |0) state, while the value | / ?|2represents the probability that the qubit will be measured in the |1) state. If more than one qubit is present, two or more qubits may be entangled. Entanglement means that the state of one qubit is dependent on the state of at least one other qubit and vice versa, wherein further in the entangled state the respective qubits cannot be regarded as individual qubits anymore. Generally, a register of N qubits in a quantum computer can be put into a superposition of basis states at once whereas a register of N classical bits can only be in a single basis state at once. Thus, in contrast to classical computing devices on a quantum computing device 2Nbasis states can be manipulated and processed simultaneously allowing for exponential intrinsic parallelism.

[0057] To perform operations on the quantum computing device the computational method to solve a given problem may be translated into qubit manipulations, which may be translated into control signals for manipulating qubits. Representations of the machine-readable instructions may include common quantum mechanical representations of operations in the Hilbert space. Depending on a specific realization of the quantum computer different representations of the qubit states may be chosen. Any state preparation on the quantum computing device may be represented by a manipulation acting on the qubit states. A manipulation may be translated into control signals to control a respective part of the quantum computer, which depend on the type of quantum computing device used. This way based on the manipulation acting on the qubit states, operations may be performed on the quantum equivalent of a classical processor as part of the quantum computing device.

[0058] In gate-based quantum computer systems the manipulations acting on the qubit states may generally be one- or multi-qubit operations. A one-qubit operation may change the state of one qubit e.g., into a specific superposition which corresponds to a rotation of the vector |X) as illustrated in Fig 1. For example, in a superconducting quantum computer this can be accomplished by microwave pulses or in a trapped-ion quantum computer by irradiation of the ion with a laser beam. A multi-qubit operation may create entanglement between two BASF SE 230213 or more qubits. For example, in a superconducting quantum computer this may be achieved by connecting qubits via an intermediate electrical coupling circuit or in a trapped-ion quantum computer via controlling the collective vibrations of the trapped ions.

[0059] Generally, to prepare manipulations for solving a given problem a respective quantum mechanical representation of the problem may be translated into qubit manipulations, which are carried out to prepare a solution of the given problem. After the preparation of the predetermined solution, i.e. after the application of the operations to the qubits of the quantum computer, a projective measurement of all individual qubits is carried out returning either 0 or 1 for each qubit. This projection usually happens in the azeigenbasis of the qubits which is also used to define the computational basis stats “0” and “1 ” of the qubit. This means that only operators that are products of azPauli operators or can be directly transformed into such operators can be measured concurrently. On the quantum computing device this measurement is achieved by applying a hardware-specific readout protocol of a series of readout manipulations including control pulses and monitoring the response to control pulses. For example, a superconducting qubit may be coupled to a hardware resonator. The measured shift of the resonator frequency allows to determine the state of the qubit as this shift depends on the state of the coupled qubit. In case of trapped ions, for example, an optical readout may be used, e.g. the state of the qubit is 1 if the ion emits light or 0 if the ion does not emit light or vice versa. This way qubits may be used to implement logical circuits or gates as in classical computing devices.

[0060] In Fig. 2 a schematic example of a quantum computer is illustrated. The quantum computing device 100 shown in Fig. 2 includes a quantum register 104 configured to perform the quantum computation, a manipulation part 106 configured to manipulate the quantum register, in particular, quantum elements forming the qubits, and a readout part 108 configured to collect measurement signals from the quantum register 104 for reading out the qubits after a quantum mechanical calculation. The manipulation part 106, in particular, provides manipulation signals for manipulating the quantum register, wherein the manipulation signals are generated based on received control signals that are determined based on the respective operations that should be performed on the qubits. In some embodiment a feedback loop between the manipulation part 106 and measurement part 108 can be provided. In contrast to classical computing, where one measurement cycle provides the state of a transistor, quantum computing includes performing multiple measurement cycles to provide a probability density or a probability for the qubit states in case of gate-based quantum computers. BASF SE 230213

[0061] The quantum register 104 can be based on different quantum elements representing the qubits. In some embodiments of gate-based quantum computers the qubits may be implemented by photons as quantum elements. Such optical quantum computing devices may include lasers that generate photons that are provided to a waveguide. A beam splitter can be provided for manipulating the photon states based on manipulation signals such as a mechanical rotation applied to a mirror. The measurement part 108 can in such an embodiment be a photon detector, and the measurement signals can be photons.

[0062] In other embodiments of gate-based quantum computers the qubits can be implemented by electronic states of ions trapped in a magnetic field. The manipulation part 106 can in such a case utilize a laser, and the manipulation signals can cause the providing of control laser pulses. Moreover, in this case, the readout part 108 can be a photon detector combined with read-out laser pulses, and the measurement signals 102 may be photons. Other qubit implementations may be based on superconductors as quantum elements, semiconducting material with anyons as quantum elements, or the like.

[0063] Fig. 3 illustrates a schematic exemplary method for generating a control signal to perform manipulations on the quantum computing device and for processing measurement signals from the quantum computing device. In most embodiments of quantum computing devices known to date, the control signals for the quantum computing device are prepared on a classical computing device and the measurement signals provided by the quantum computing device are further processed on the classical computing device. Other embodiments are, however, conceivable as quantum computing devices mature. In the following example, the quantum computer refers to a gate-based quantum computer for which the manipulations refer to operations on the quantum elements of the quantum computer.

[0064] For generating the control signal to perform operations on the quantum computing device, the problem to be solved with the aid of the quantum computing device is provided in step S10, preferably, in a mathematical description. Such problem may for instance include an optimization problem for training a data-driven model based on a training data set. For example the problem can include an optimization problem associated objective functions being optimized during the training performed by the quantum computer. Based on the problem to be solved, an operation description of the problem or a sub-problem may be generated in step S12, wherein the operation description comprises the operations to be applied to the qubits of the quantum computer to solve the problem in the quantum mechanical calculation. Further, the operation description can include a reference state that allows to generate a representation of an initial qubit state on the quantum computer on which the further operations are then applied by manipulating the qubit states. Based on BASF SE 230213 the operation description control signals can then be generated in step S14 to control the quantum computer, for instance, by providing the control signals to the manipulation unit that can then manipulate the qubit states based on the control signals. In step S16 the manipulation unit then applies the manipulation operations to individual or multiple qubits of the quantum computer, wherein based on the manipulation operations the qubits perform the quantum mechanical calculation. After the manipulation, measurement signals can be generated to determine the result of the quantum mechanical calculation in step S18. This step can include a read-out, i.e. measurement, of the qubit states after applying the manipulation operations to the initial qubit states. The measurement signals can in step S20 then be translated into a measured quantity on the classical computer and in case of a subproblem fed back into the problem to be solved. Finally, the result of the problem calculation including the quantum mechanical calculation can be provided on the classical computing device in step S22.

[0065] Fig. 4 illustrates a schematic example of a hybrid system including a classical and a quantum computing device. As described with respect to the method illustrated in Fig. 3, quantum computing devices are often used in connection with classical computing devices. As shown in Fig. 4 a problem preparation system, e.g. control signal generation apparatus, can be realized as a classical computing device 110 performing, for instance, steps S10, S12, S20, S22 of the method illustrated in Fig. 3. A controlling unit can then be provided as interface between the classical computing device 110 and the quantum computer 100, wherein the controlling unit can also be a classical computing device, for instance, performing step S14. The control unit can then be communicatively coupled with the manipulation part 106 that can control the manipulators of the quantum computing device. Also, the manipulation part 106 can be realized as a classical computing device, for instance, a classical controlling hardware for the control of specific hardware components of the quantum computer that perform the manipulation of the qubits. However, the manipulation part 106 is generally regarded as part of the quantum computer, since it directly influences the quantum register. The quantum computing device 100 is adapted to perform the quantum operation S16, in particular, by the manipulation of the qubits of the quantum register. The measurement part 108 that is also generally regarded as part of the quantum computing device can then perform the step S18 by utilizing classical hardware. The measurement part 108 can then be communicatively coupled to the preparation system 110 for further processing of the measurement signals.

[0066] Fig. 5 illustrates a schematic example of a quantum computing device based on superconductors. Superconducting quantum computing devices are one of the solid-state quantum BASF SE 230213 computing technologies. Here the quantum register 104 can include superconducting circuits 520, 522, 524 based on Josephson junctions. The qubits can then, for instance, refer to charge, flux, transmon, or phase qubits depending on the quantity of the superconducting circuits that are chosen to represent the qubits. Fig. 5 refers to a simplified illustration of a superconducting quantum computer utilizing charge qubits. For charge qubits the different states of the qubit are represented by an integer number of Cooper pairs on a superconducting island. In case of gate-based quantum computing quantum manipulations can then be performed by manipulating the qubits through microwave pulses. Resonators 512, 514, 516 can be utilized to manipulate the state of the qubits by applying the microwaves or for reading out the state of the qubits by measuring respective microwaves, wherein generally different resonators are used for the manipulation of the state of the qubits and the readout of the qubits. Moreover, resonator 518 can be utilized for applying microwaves that entangle the qubits. However, instead of resonator 518 the entanglement can also be achieved by an inductive or capacitive coupling of the superconducting circuits or even by providing another qubit, here a superconducting circuit, between the to be entangled qubits.

[0067] On an operational level such systems are maintained at extremely low temperatures, e.g., in the tens of mK. The extreme cooling of the systems keeps superconducting materials below their critical temperature and helps to avoid unwanted state transitions. To maintain such low temperatures, the quantum information processing systems may be operated within a cryostat, such as a dilution refrigerator. In some implementations, control signals are generated in higher-temperature environments, and are transmitted to the quantum computer using shielded impedance-controlled GHz capable transmission lines, such as coaxial cables. In some implementations, the state measurement of superconducting qubits is achieved using a dispersive detection scheme. In order to read out or detect the state of any qubit, a probing signal, e.g., a travelling microwave, may be excited along a readout transmission line coupled to the qubit via a respective readout resonator. The frequency of the probing signal can be in the vicinity of the resonance frequency of the readout resonator. Depending on the internal quantum mechanical state of the qubit, the intensity or phase of the probing signal transmitted along the readout transmission line may be altered because the reflectivity of the readout resonator coupled to the qubit changes depending on the state of the qubit. This allows for the state detection of the qubits, wherein during the readout of a qubit state the state of the qubit collapses, i.e. is projected with the respective probability onto one of the basis states. By performing the quantum mechanical calculation and the readout a plurality of times the respective probabilities can be determined. Further details for superconducting quantum devices are described e.g. in documents EP 3830867 A1 , EP 3449427 A1 , US 2020272925 A1 , CN 212061223 U and US 2019019099 A1. BASF SE 230213

[0068] Fig. 6 illustrates a schematic example of a quantum computing device based on ions in an ion trap. Similar to neutral atom traps ion traps with, e.g. positively charged Calcium ions, can be used to implement the quantum computing device. Here ions 626 are trapped in an oscillating electromagnetic field 624 inside a high or ultra-high vacuum. The ions 626 are laser cooled and held in the oscillating electrical field 624. For qubit manipulation such as superposition or entanglement laser light 628 at different frequencies may be used.

[0069] Generally, based on the above described quantum computer realizations gate-based type calculations can be performed on a quantum computer hardware architecture. The gatebased type calculation is based on quantum gates. In contrast to classical gates, there is an infinite number of possible single-qubit quantum gates that can change the state vector of a qubit. Changing the state of a qubit state vector typically is referred to as a single qubit rotation, and may also be referred to herein as a state change or a single-qubit quantum gate operation. A rotation, state change, or single-qubit quantum gate operation can be represented mathematically by a unitary 2 x 2 matrix with complex elements. A rotation corresponds to a rotation of a qubit state within its Hilbert space, which can be conceptualized as a rotation of a vector on the Bloch sphere, wherein the Bloch sphere is generally known as a geometrical representation of the space of the pure states of a qubit. Multiqubit gates alter the quantum state of a set of qubits. For example, two-qubit gates rotate the state of two qubits as a rotation in the four-dimensional Hilbert space of the two qubits, wherein, as generally known, the Hilbert space is an abstract vector space possessing the structure of an inner product that allows length and angle to be measured. Furthermore, Hilbert spaces are complete, i.e. there are enough limits in the space to allow the techniques of calculus to be used.

[0070] Not all quantum computers are gate-based quantum computers. Embodiments of the present invention are not limited to utilizing gate-based quantum computers. As an alternative and in particular advantageous example, embodiments of the present invention can also utilize, in whole or in part, a quantum computer that is implemented using a quantum annealing paradigm which is an alternative to the gate-based quantum computing paradigm. More specifically, quantum annealing is a metaheuristic for finding the global minimum of a given objective function over a given set of candidate solutions (candidate states), by a process using quantum fluctuations. In particular, quantum annealing is closely related to adiabatic quantum computing.

[0071] In more detail, quantum annealing is a heuristic quantum optimization algorithm which works on the basis of adiabatic quantum computation governed by an adiabatic process. | BASF SE | 230213

[0072] Evolution of a quantum system defined by a state is governed by Schrodinger’s Equation, as shown below. where i is the unit imaginary number, h is the Planck’s constant, t is time, and J£(t) is the Hamiltonian of the system. An adiabatic process is one for which the initial Hamiltonian J£initchanges very slowly to some different final Hamiltonian J£final. The adiabatic theorem states that if a system is in the nth energy state (or nth eigen vector) of J£init, then the system will be in the nth energy state of J£finalunder adiabatic change of the Hamiltonian. The preparation of the algorithm requires the J£initand J£final, and at any point of time t, the Hamiltonian is described by J£(t), where,

[0073] Wherein T is the time over which the adiabatic change should occur. Quantum Annealing is a process of slowly changing J£(t) from the ground state of a known initial state of J£initto the problem Hamiltonian J£final. ^mit is typically chosen such that it refers to an easy-to- prepare state on the quantum annealing hardware. In case of an optimization problem, like the training of a data-driven model J£finalcan be chosen based on the Ising model resulting in a quadratic unconstrained binary optimization problem that can be solved by the quantum computer as described in more detail in the following sections. For a system to stay in its ground state, the time T over which the adiabatic change should occur is given by where AAtis the difference between the first excited energy and the ground energy of J£(t).

[0074] A quantum annealer is in particular suitable for solving complex optimization problems and thus can in an embodiment of the invention advantageously be utilized. Quantum annealing works by minimizing a system of interconnected qubits to their lowest energy state. The system is initially set as a quantum superposition of states, where there is an equal probability for all spin permutations of the qubits (all possible solutions to the problem). At the end of quantum annealing the qubits either end up in a spin-up (1), or spin-down (0) state which encodes the solution to the respective optimization problem. The problem solution can then be achieved by the slow and smooth evolution of the spin system to its ground state referring to the optimal solution to the optimization problem. BASF SE 230213

[0075] Generally, also quantum annealing procedures start with utilizing a classical computer providing or generating an initial Hamiltonian and a final Hamiltonian based on a computational problem to be solved, and providing the initial Hamiltonian, the final Hamiltonian and an annealing schedule as input to a quantum computer. In case of an annealing procedure in which an optimization problem is solved, preferably, the final Hamiltonian refers to an Ising Hamiltonian representing the optimization problem. The quantum computer is then adapted, for instance, by utilizing a respective control unit controlling a manipulation part of the quantum computer, to prepare a relatively easy to prepare initial state, such as a quantum-mechanical superposition of all possible states, e.g. candidate states, with equal weights, based on the initial Hamiltonian. After the preparation of the initial state on the quantum computer, the initial state is then evolved according to the annealing schedule following a time-dependent Schrodinger equation referring to a natural quantum-mechanical evolution of the physical system of the quantum computer. More specifically, the state of the quantum computer undergoes time evolution under a time-dependent Hamiltonian, which starts from the initial Hamiltonian and terminates at the final Hamiltonian. If the evolution rate is slow enough, the system stays close to the ground state of the instantaneous Hamiltonian. At the end of the time evolution, the set of qubits, i.e. quantum elements, on the quantum annealer is in a final state, which is expected to be close to the ground state of the Ising Hamiltonian that corresponds to a solution to the original problem, referring, for instance, to an optimization problem. The final state of the quantum computer can then be measured, thereby producing results that can be utilized for solving the original problem. The measurement operation can be performed, for example, in any of the ways described already above. A classical computer can then perform postprocessing on the measurement results to produce an output representing a solution to the original computational problem. A quantum annealer as described above can, for instance, be realized on a superconducting quantum computer hardware.

[0076] Fig. 7 shows schematically and exemplarily an example of hardware principles and structures of a quantum annealer that can be advantageously utilized as quantum computing unit in embodiments of the invention. In this example, the quantum annealer hardware comprises a classical computer part and a quantum elements part. The classical computing part is configured to interact with respective quantum elements of the quantum computing part, for example, for controlling, manipulating or measuring respective quantum elements as described above. Further, the quantum annealer hardware comprises an interface to the classical computer that allows to input a problem and to output a respective solution of the problem or at least respective measurement results of the computation of the input problem. The classical computer of the classical computing part is configured to control the operation of the respective quantum elements of the quantum part. However, the classical BASF SE 230213 computer can also be configured for additional tasks. For example, the classical computer can be configured to provide a problem translation of a provided problem into a respective quantum mechanical description or quantum mechanical algorithm that can be performed on the respective specific quantum element hardware. Moreover, the classical computer allows to translate the specified quantum algorithm and operations included in the algorithm into the specific control signals that allow to manipulate the quantum elements. Since each quantum computer can have different manipulation and controlling mechanisms built into the hardware of the quantum computer, this translation step can be very specific for each quantum computer and is thus in most cases provided by a classical computer specifically configured for this task for a specific quantum computer. Further, the classical computer allows to control the measurements of respective quantum elements as a result of a quantum computation and also a translation of the measurements, for instance, of respectively measured states of the quantum elements into electronically and digitally utilizable quantities. Further, the translation can also refer to providing an overall result of the quantum computation which might include averaging, error correction, etc.

[0077] The quantum element part of the quantum annealer hardware comprises the quantum annealer itself and optionally an additional quantum governor. The optional quantum governor can be considered as a class of non-information bearing degrees of freedom that can be configured to steer a dissipation dynamics of information bearing degrees of freedom of the quantum annealer. For example, the quantum governor can be configured to navigate the quantum evolution of a disordered quantum annealing hardware at finite temperature in a controlled manner and improve the adiabatic quantum computation process. This can be achieved, for example, by the quantum governor by facilitating a driving of the quantum annealer towards a quantum phase transition while at the same time decoupling the quantum annealer from excited states by making excited states effectively inaccessible by the quantum annealer. The quantum annealer itself can be realized, for instance, by a programmable quantum chip. For example, such a quantum chip can include a predetermined number of cells comprising a respectively predetermined number of qubits. The qubits can be connected by programmable inductive couplers, wherein the qubits of a cell are connected to each other but can also be connected with the qubits of other cells in specific predetermined patterns that define the programmability of a specific quantum annealer. A particular advantageous example of a quantum annealer that can be utilized in this invention can be found, for instance, in the document EP 3 092 607 A1 incorporated herein by reference.

[0078] In the following the term operation description refers to a representation of a problem that comprises a sequence of quantum operations that should be applied during a quantum BASF SE 230213 mechanical calculation of the problem. The term “quantum operation” can include in the context of this invention all types of quantum gates as described above and more generally all manipulations known for quantum elements on any quantum computer hardware. Moreover, the term can also include operations performed on components of the quantum computer representing a coupling between the quantum elements forming the qubits. These operations then refer to any kind of change of the state of the coupling representing components, for example, a turning of a coupling on and off, or the change of a field frequency, etc. Further, in some applications the quantum operations can also include measurement operations. This allows to implement algorithms using a measurement feedback. For example, in such an algorithm a quantum computer can execute the quantum gates defined by the sequence of quantum operations and then measure only a subset, i.e., fewer than all, of the qubits in the quantum computer, and then decide which further quantum manipulations to execute next based on the outcome of the one or more measurements. In particular, measurement feedback can be useful for performing quantum error correction, but is not limited to use in performing quantum error correction.

[0079] Fig. 8 shows schematically and exemplarily an architecture of an asset transaction system controlled and / or monitored by an apparatus for controlling and / or monitoring an asset transaction according to the invention. In the example shown in Fig. 8, the assets are transferred utilizing a distributed ledger technology. The distributed ledger technology can validate, store and link respective transactions of digital representations of assets, for instance, in blockchains. The ledger of respective assets are stored and maintained by a plurality of ledger nodes, wherein users, for instance, owners of assets, can participate in the asset transfer utilizing the ledger by utilizing respective wallets as interface to the ledger nodes. The wallets can be utilized to send and retrieve respective asset transactions to and from the ledger node network. In this context, the invention provides a control and / or monitoring system that can interact with the ledger and / or the wallet, for instance, to monitor the transactions performed in the ledger or to control transactions performed by the wallet. A respective exemplary workflow for a controller and / or asset monitoring system is described in the following.

[0080] Fig. 9 shows a general concept of the invention. In this context, historical asset data is utilized and processed to determine respective technical indicators, for instance, a feature set that is indicative of the historical asset data. For example, the historical asset data can refer to time series monitoring data monitoring one or more aspects of asset transfers. Further, from the historical asset data also historically performed asset transfers or monitoring events can be derived. A respective data-driven model can then be utilized to learn based on the technical indicators, for instance, respective feature sets representing the BASF SE 230213 historical asset data, when and how asset transfers have been performed or monitoring events have occurred. This allows the data-driven model in a next step also to control or monitor asset transfers in real-time based on respective current real-time asset monitoring data.

[0081] Fig. 10 shows schematically and exemplarily a workflow that can be performed by a controlling and / or monitoring system for controlling and / or monitoring one or more asset transfers. The controlling and / or monitoring system can be realized in form of one or more classical computers and at least one quantum computer unit. In the workflow, first time series monitoring data related to the asset for which the transfer should be controlled and / or monitored can be provided, for instance, via a respective interface. The time series monitoring data can be any data related to the asset and providing information that can be relevant for an asset transfer. For example, the time series monitoring data can relate to a transfer value of a respective asset transfer. However, the time series monitoring data related to the asset can also refer to measurement data, for instance, related to a CO2 measurement. Depending on the respective source, the interface can be configured to receive the respective time series monitoring data and provide the monitoring data, for instance, to an interface of a training system for training a data-driven model that can be utilized for controlling and / or monitoring one or more asset transfers. The monitoring data can additionally or alternatively also be provided to a classical computer utilizing the monitoring data directly for controlling and / or monitoring one or more asset transfers. If the monitoring data is utilized for training, the monitoring data can be stored, for instance, as historical monitoring data and can then be processed to generate and provide respective training data for the training of a data-driven model. However, the historical monitoring data can also be received and processed from other sources.

[0082] A respective training data set can be provided via a respective interface to a classical computer for training a data-driven model. A training data set includes, a) respective control and / or monitoring data of asset transfers and b) time series monitoring data related to the asset. For example, the time series monitoring data related to the asset can be regarded as features in the context of a training of the data-driven model and the control and / or monitoring data that is associated with asset transfers and the time series monitoring data can be regarded as a label to the features such that the training data set refers to a labelled data set fortraining. For example, the control and / or monitoring data of such asset transfers can be associated with respective time series monitoring data by unambiguously relating the timing of asset transfers and / or monitoring events and the respective control and / or monitoring data utilized during the asset transfers with the time series monitoring data at BASF SE 230213 that time. To train a data-driven model that is configured to provide a model output indicative of control and / or monitoring data for controlling and / or monitoring an asset transfer when provided with time series monitoring data as input, the training data set can then be provided by the interface to a classical computer. The classical computer can then be configured to initialize the data-driven model based on the training data set. Details and examples with respect to such an initialization will be provided with respect to specific examples below. Based on the initialized model, the classical computer can then be configured to generate control data for controlling a quantum computer to determine discrete weights of the data-driven model based on the initialized model and the training data set. For example, the initialized model can refer to an optimization function that when optimized based on the training data set on the quantum computer provides respective discrete weights of the data- driven model. The discrete weights of the data-driven model can be regarded as comprising at least partly the information on the functional relation between the time series monitoring data and the control and / or monitoring data of asset transfers that has been learned by the data-driven model.

[0083] The respectively trained data-driven model can then, for instance, be stored on a respective storage unit or can be directly provided to a classical computer performing a workflow for controlling and / or monitoring one or more asset transfers in real-time. For this, the classical computer can be configured in an optional step to aggregate the monitoring data, wherein the aggregation can include determining a feature set representing the time series monitoring data for a predetermined time interval. The features determined during the aggregation step can strongly depend on the respective application case, for instance, on the respective asset and the respective time series monitoring data related to the asset that are utilized. Examples for the respective feature set will be provided in following more detailed examples. In a next step, a trained data-driven model can be provided, for instance, by accessing a storage on which the respective trained data-driven model has been stored or by directly receiving the trained data-driven model from the training apparatus described above. The aggregated data can then be provided as input to the data-driven model, wherein the data-driven model then processes the input to generate a model output that is indicative of control and / or monitoring data. For example, the model output can refer to discrete quantities, for instance, “0” and “1 ”, wherein the output “0” can then be indicative of not transferring an asset, wherein the output “1 ” can be indicative of generating control data for transferring the asset. Thus, based on the model output, the respective control and / or monitoring data can be generated. The generated control and / or monitoring data can then be utilized, for instance, provided to respective interface for asset transfer like a wallet, for controlling and / or monitoring a respective transfer of the asset. BASF SE 230213

[0084] In the following, some particularly advantageous and more detailed embodiments of the invention will be described. In an advantageous embodiment the data-driven model is based on a boosting algorithm comprising a plurality of weak classifiers. The data-driven model can then be trained based on an iterative machine learning algorithm that solves a hard optimization problem in each iteration to select a subset of weak classifiers from a larger set. Thus, the training leads to a construction of a strong classifier as data-driven model by concatenating the subsets of weak learners selected in each iteration.

[0085] Fig. 11 outlines an example of principles of training a boosting algorithm that can be utilized as data-driven model in the invention. The training can be performed in two steps. In a first step a plurality of predetermined weak learners are trained on a classical computer using a variant of boosted training as part of initializing the data-driven model. In a second step a quantum boosting optimization problem can constructed based on the trained weak classifiers as initialized model that can be solved using a quantum computer. The main objective of the quantum boosting optimization is to reduce the number of weak learners in the final ensemble of the data-driven model by reducing the correlation between the weak learners in the final ensemble.

[0086] T raining the weak learners in the first step as part of the initialization of the model on a classical computer can comprise the following steps schematically shown in the right part of Fig. 11 . Starting from the same weight for all the data samples derived from the training data, where each sample has the index s and the data vector xs, and initializing the first weak learner D±with a uniform distribution s such that £>i(s) = 1 / S where S is the total number of data samples, for every sequentially appended weak learner ht, i = 1,2, .... N the following quantity can be determined with: where ysis the label of the sth data sample. The sample weights can then be defined as wherein E is a variable indicating the learning rate, and the weight distribution can be updated in each iteration step of the training process as follows with a normalization factor Ztdefined such that Di+1is a probability distribution. This equation can be regarded as part of an initialized model in this embodiment. Dtis then used to train the weak learner h by acting as a weight in the sum for the calculation of the loss BASF SE 230213 function. For example, if the loss function being used is the mean squared error then the value of the loss function for the weak learner would be:

[0087] In a next step the initialized model comprising results of the trained weak learners as above can be provided to a quantum computer for training the discrete weights btusing the quantum hardware. The quantum trained weights of the data-driven model are discrete weights. For example, the discrete weights can be binary weights of 1 or 0 to determine if a weak learner is part of the ensemble of the final data-driven model. To map the objective function onto the quantum hardware the weak learner ensemble weights in the initialized model, can be encoded as binary variables which can be written as a vector of binary or discrete weights, b. The initialized model can then be trained utilizing a learning objective function, for example, given by where L(b) is the loss function of the classification, R (b) is the regularization term, parameterized by A, in {-1, +1} is the classification of the ith weak learner, ysis the label of the sth data sample and xsis the sth input data sample. A key advantage of using a quantum formulation is that the discrete LO-norm in R (b) , can be used. The LO-norm counts the number of active classifiers in the ensemble and has been shown to have advantages over other measures commonly used. By optimizing the weights b using the objective function H0(b) a set of optimal binary weights boptis obtained.

[0088] After the training, utilizing the above optimization given a new data point x from the time series monitoring data, the final strong classifier as data-driven model can be constructed with the selected classifiers from the optimization procedure. To obtain a classification C( x) for x, the following equation can be used:

[0089] C(x) = sign where T is an optional parameter that can either be set as 0 or as a predetermined threshold that can enhance the results and can be computed in a post-processing step with: BASF SE 230213

[0090] The challenges of the training of a data-driven model as outlined above lie in this embodiment in the optimization problem to be solved. In particular, the loss function that estimates the error that a classifier should cause over a set of examples has a strong influence on the computational resources necessary for solving the problem. The most advantageous choice is an LO-norm penalization since it brings the weights of the undesirable classifiers to exactly 0. However this leads to a NP-hard optimization problem due to non-convexity. Instead of approximating this problem on a classical computer, the quantum computer allows to solve this problem, as described above, for real-life applications, in particular, in the context of controlling and / or monitoring asset transfers.

[0091] In the following an advantageous example of an embodiment of the invention is described. Fig. 12 illustrates an embodiment of an exemplary system for controlling and / or monitoring an asset transfer. The system comprises four sub-systems comprising a data aggregation unit, a transfer signal generation unit, a data-driven model training unit and a user interface. The units can be implemented using classical computing software and hardware, wherein the model training unit can be implemented to as classical-quantum computing hybrid. Moreover, also the transfer signal generation unit can be implemented to utilize a quantum computer for applying the data-driven model to the time series monitoring data.

[0092] The part of the above system implemented on a classical computer indicated is illustrated in more detail Fig. 13 showing an exemplary embodiment of a flow of data inside the system between respective sub-systems and units. The data aggregation unit can comprise a subunit that receives time series monitoring data from an external data provider depending on the respective application. This time series monitoring data can then be saved into a data store, for example onto a hard-drive as a data file or into a database.

[0093] The data aggregation unit can comprise a unit that is configured to aggregate the time series monitoring data. The stored time series monitoring data can be loaded and aggregated into regular, or irregular spaced time intervals. For example, time series monitoring data may be aggregated into regular space 5-minute intervals comprising respective monitoring data. The aggregated data can then be stored into either physical or virtual memory. It may also be directly passed via a data pipeline to transfer signal generation unit. Optionally, the data aggregation unit or the transfer signal generation unit can comprise a preprocessing unit that processes the aggregated data into new sets of quantitative values. These new sets of processed values can be referred to as feature sets. These feature sets can be stored into either physical or virtual memory. The feature sets can also be directly passed via a data pipeline to the computational inference unit. BASF SE 230213

[0094] The computational inference unit of the transfer signal generation unit is configured to perform a computational inference using the aggregated or optionally the feature sets as input data. The computational inference unit can be configured to utilize an asset value part of the data-driven model to generate a binary output. The input data can be processed in a time sequential manner to produce a sequence of binary outputs. In an example this sequence of binary outputs can indicate whether an asset value will increase (+1) or decrease (-1) in a respective predetermined time step. In an example, the time series monitoring data can be aggregated into regular space 5-minute intervals, wherein in this case computational inference unit can output a binary output at each 5 minute interval.

[0095] The transfer signal generation unit can further comprise a decision function unit that is configured to execute a decision function, for example, as part of a control data part of the data-driven model. The decision function can utilize the binary output as input to generate a model output indicating, for example, to perform an asset transfer to a user (+1), do not transfer an asset (0), or perform an asset transfer from the user (-1). The computer interference unit and the transfer signal generating unit are described in this example, as separate units each utilizing a part of the data-driven model. However, in other examples, the units can also be realized as one unit that utilizes the data-driven model, or the data-driven model can be realized with only on part directly determining the model output.

[0096] Fig. 15 shows schematically and exemplarily and embodiment of the model training unit. The model training unit can comprise a training data providing unit that can be configured to load historical aggregated data and generate labels of, for instance, asset transfers, based on optimal transfer decisions that can be determined based on the historical monitoring data. However, the labels can also be generated based on actual asset transfers performed during the respective time of the historical monitoring data. Moreover, the labels can also refer to monitoring events. The respective target transfer decisions can then be referred to as the training labels and can form together with the historical time series monitoring data the training data set. In the following example, it is referred to asset transfers and thus to a controlling aspect of the asset transfer, however, the same principles can also be applied to monitoring events in the monitoring aspect of the asset transfer.

[0097] The model training unit can further comprise a processing unit configured to process the historical aggregated data into feature set following the same processing as described with respect to the preprocessing unit described above. Moreover, the model training unit can comprises a training unit that trains the data-driven model, in this example, the asset value part of the data-driven model, using the training data set including the feature sets as input to determine as output a binary output. The training is performed on a quantum computer such that the binary output as closely as possible matches the training labels given at each | BASF SE | 230213 time step and when used with the decision function the binary output aim to maximize a respective objective of the asset transfer controlling, for example, transferring assets that are associated with a reduced CO2 production.

[0098] The user interface can comprise an information providing unit that may visually or otherwise indicate to a user a suggested monitoring and / or controlling action that is based on the respective monitoring and / or controlling data determined utilizing the model output.

[0099] Further exemplary more detailed embodiments of the units described above are provided in the following. In an embodiment, the data-driven model as shown in Fig. 14 can be defined as a series of aggregated computational methods that processes data and output a signed binary value, e.g. either 1 , -1 or 0, that may be used to indicate the transfer or monitoring actions for an asset. In a preprocessing step as described above, the time series monitoring data can be aggregated, for example, into feature sets comprising one or more quantitative values, that can be referred to as technical indicators, and that can be provided as input to the data-driven model. Using procedures for computational inference the data- driven model may then for example apply statistical, numerical or other computational techniques to the technical indicator values to output a signed binary value.

[0100] Technical indicators, i.e. features, may be, but are not limited to, sets of mathematical functions that are suitable for a respective application. For example the feature sets can comprise at least one of a simple moving average (SMA), a relative strength index (RSI), a rate of Change (ROC) and a moving average converge-divergence (MACD). The computational inference described above comprises using as part of the data-driven model a machine learning algorithm that can maps the input feature sets, which may be technical indicators, to respective model output indicative of transfer and / or monitoring actions. In an embodiment, the machine learning algorithm as part of the data-driven model can referred to a boosting algorithm, for example, as previously described above.

[0101] Using the technical indicators, i.e. features, of the feature sets as inputs to the machine learning algorithm of the data-driven model a model output is generated per time step. For example, using a boosting algorithm as described above the binary output, C(xt), at timestep t, with the technical indicators presented as vector xtis calculated as | BASF SE | 230213 where wtis the weight of the ith classifier, h is the binary output of the ith classifier, T is a threshold value and N is the total number of classifiers. The decision function as part of the data-driven model can be a function that takes as inputs the binary output from the computational inference and outputs a respective model output utilized for generating controlling and / or monitoring data, for instance, whether an asset should be transferred or not. For example, the decision function for controlling an asset transfer can follow, where the model output can be generated when the binary output changes direction, otherwise a 0 value indicating to not transfer an asset can be generated. The output of the decision function can then be utilized to generate the monitoring and / or controlling data and, for instance, suggest a respective action to a user via the user interface.

[0102] The data-driven model as described above can be trained at least partly using a hybrid quantum-classical paradigm. An example of an embodiment of a training workflow is shown in Figure 15. In a first step historical aggregated time series monitoring data can be loaded from a data storage system. The historical aggregated time series monitoring data can then be prepared for training. For example, the historical aggregated time series monitoring data can be processed using the same methodology as described for determining the feature sets above. This may comprise processing the historical aggregated time series monitoring data into a set of technical indicators as discussed previously. Further, labels can be generated from the raw time series monitored data rt. For example, optimal direction labels, yt, can be generated at each time step with respect to an application objective. These labels can be defined as either +1 indicating that an asset value increases in a predetermined time step or -1 if the asset value decreases in the predetermined time step. The predetermined time step can be a past, present and also future time step. A non-negative threshold value can also be applied, wherein if the determined asset value is not above or below this threshold then the label from the previous timestep is used. if — yt> threshold, yt= if yt> threshold, if l / tl — threshold, where, where ytis the respective target application objective also known as the training target, and the quantity IE [rt+T] is the expected asset value for the predetermined time step. The expected asset value quantity may be directly read from the time series, by looking at the data T steps ahead or by other numerical methods such as Monte Carlo simulations. | BASF SE | 230213

[0103] For example, if T = 2, where an asset value is evaluated 2 timesteps ahead with a threshold of 0, generated labels are shown in the table below:

[0104] The historical aggregated time series monitoring data and the corresponding labels can then form the training data set. The training data set can then be split into two datasets with a ratio of size such as 3:1 . The first set of data can then be used as input to the training procedure. The preprocessed time series monitoring data of the first set is used as model inputs and the targets for the model to learn are the corresponding optimal direction labels. The second set of data can be referred to as the validation data. The validation data can be independent of the first data set used for training and is used to evaluate the model performance on unseen data. Once the respective data sets have been generated the models can be trained with the objective of learning to take the processed time series monitoring data as input and output the correct label, yt, at each time step.

[0105] Multiple models may be trained either sequentially or in parallel. For each model different model parameter settings can be used. This is to enable post selection of the best using a validation data set. The utilized model parameters may be, a number of weak learners, a depth of the decision trees of the weak learners, other relevant parameters of weak learning models, a regularization parameter, etc. For each of the models the set of parameters can be selected from a discrete grid of parameter combinations or can be randomly sampled. After the model parameters have been chosen the respective model can be trained.

[0106] The training in this example can comprise the following steps. In a first step for initializing the model the weak learners can be trained on a classical computer using boosting or other such methodologies. The outputs of the final trained weak learners can then be used to formulate the quantum boosting objective function as part of the initialized model. This objective function aims to find the best set of weights for the weak learners to combine their outputs in order to minimize the error of the combined output direction label with respect to the known label at the time step. For example, it is advantageous to formulate the objective function as quadratic-unconstrained binary optimization (QUBO). The QUBO can then be sent as part of the initialized model to the quantum computer to be optimized using for BASF SE 230213 example quantum annealing hardware, or variational algorithms on gate based hardware. The output of the optimal solutions then refers to a list of normalized weights for each weak learners output. Some weights may be set to 0 which excludes the given weak learner from the final model. The optimization can also be configured to reduce the complexity of the final model by reducing the number of non-zero weights.

[0107] The final, i.e. trained, model, given by the combined weak learners, can then be evaluated on another dataset, with data that has not been used in the training, known as the validation data. The model direction label predictions are evaluated on the validation data by calculating respective performance metrics such as a machine learning F1 score. This metric can be important because it indicates the model's ability to accurately identify both positive and negative transfers. Further a respective application metric can be determined. For example, the model can be run through back-testing where following the steps of the monitoring and / or controlling data generation as described above the model outputs are transformed into controlling and / or monitoring data. Using these controlling and / or monitoring data on the time series monitoring data application metrics can be calculated. For instance, it can be determined if the model allows to meet additional target application objectives. Moreover, the models can be evaluated based on the metrics calculated from their performance on the validation data set.

[0108] The metrics can then be ranked according to their level of importance for a respective application. For example, the ranking can follow this order with the highest priority given to the a fraction score measuring an accuracy of the model's determined transfer and / or monitoring actions compared to the actual transfer and / or monitoring actions. For example the fraction score can be calculated as the ratio of the number of determined transfer actions to the actual number of transfers. In second place the target application objective can be utilized, that depends on the respective application and third can be the F1 score. The best ranked model can then be stored in the data storage to be used for computational inference on real live data.

[0109] Due to the current limitation of quantum computing hardware the invention as described above uses the quantum software and hardware in an offline framework. Herein, offline, refers to the execution of software on a classical and / or quantum device or the execution of a hardware device outside the area of the critical execution of the application. Critical execution is defined as the execution of software or hardware on a classical and / or quantum computing system that needs to be successfully executed in accordance with constraints such as, but not limited to: computational time and numerical accuracy of output. The area of critical execution can refer to the execution of the controlling and / or monitoring data generating system during the interval of receiving new time series monitoring data BASF SE 230213 and / or updating the user interface and / or implementing the respective new controlling and / or monitoring data. The controlling and / or monitoring actions caused by the controlling and / or monitoring data may need to be generated as quickly and as accurately as possible depending on the respective application. The execution of software in the area of critical execution is executed using classical computing software and hardware for instance as shown in the left hand box of Fig. 12. Generally, the system parts shown on the left side and the right side of Fig. 12 can be run independently of each other, wherein only the part on the right side utilizes quantum hardware.

[0110] Other variations to the disclosed embodiments can be understood and effected by those skilled in the art in practicing the claimed invention, from a study of the drawings, the disclosure, and the appended claims.

[0111] For the processes and methods disclosed herein, the operations performed in the processes and methods may be implemented in differing order. Furthermore, the outlined operations are only provided as examples, and some of the operations may be optional, combined into fewer steps and operations, supplemented with further operations, or expanded into additional operations without detracting from the essence of the disclosed embodiments.

[0112] In the claims, the word "comprising" does not exclude other elements or steps, and the indefinite article "a" or "an" does not exclude a plurality.

[0113] A single unit or device may fulfill the functions of several items recited in the claims. The mere fact that certain measures are recited in mutually different dependent claims does not indicate that a combination of these measures cannot be used to advantage.

[0114] Procedures like the providing of the time series monitoring data, the providing of the trained data-driven model, the generating of the control and / or monitoring data, providing of the control and / or monitoring data, etc. performed by one or several units or devices can be performed by any other number of units or devices. These procedures can be implemented as program code means of a computer program and / or as dedicated hardware.

[0115] A computer program product may be stored / distributed on a suitable medium, such as an optical storage medium or a solid-state medium, supplied together with or as part of other hardware, but may also be distributed in other forms, such as via the Internet or other wired or wireless telecommunication systems. BASF SE 230213

[0116] Any units described herein may be processing units that are part of a classical computing system. Processing units may include a general-purpose processor and may also include a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), or any other specialized circuit. Any memory may be a physical system memory, which may be volatile, non-volatile, or some combination of the two. The term “memory” may include any computer-readable storage media such as a non-volatile mass storage. If the computing system is distributed, the processing and / or memory capability may be distributed as well. The computing system may include multiple structures as “executable components”. The term “executable component” is a structure well understood in the field of computing as being a structure that can be software, hardware, or a combination thereof. For instance, when implemented in software, one of ordinary skill in the art would understand that the structure of an executable component may include software objects, routines, methods, and so forth, that may be executed on the computing system. This may include both an executable component in the heap of a computing system, or on computer- readable storage media. The structure of the executable component may exist on a computer-readable medium such that, when interpreted by one or more processors of a computing system, e.g., by a processor thread, the computing system is caused to perform a function. Such structure may be computer readable directly by the processors, for instance, as is the case if the executable component were binary, or it may be structured to be interpretable and / or compiled, for instance, whether in a single stage or in multiple stages, so as to generate such binary that is directly interpretable by the processors. In other instances, structures may be hard coded or hard-wired logic gates, that are implemented exclusively or near-exclusively in hardware, such as within a field programmable gate array (FPGA), an application specific integrated circuit (ASIC), or any other specialized circuit. Accordingly, the term “executable component” is a term for a structure that is well understood by those of ordinary skill in the art of computing, whether implemented in software, hardware, or a combination. Any embodiments herein are described with reference to acts that are performed by one or more processing units of the computing system. If such acts are implemented in software, one or more processors direct the operation of the computing system in response to having executed computer-executable instructions that constitute an executable component. Computing system may also contain communication channels that allow the computing system to communicate with other computing systems over, for example, network. A “network” is defined as one or more data links that enable the transport of electronic data between computing systems and / or modules and / or other electronic devices. When information is transferred or provided over a network or another communications connection, for example, either hardwired, wireless, or a combination of hardwired or wireless, to a computing system, the computing system properly views the connection as a transmission medium. T ransmission media can include a network and / or data links which BASF SE 230213 can be used to carry desired program code means in the form of computer-executable instructions or data structures and which can be accessed by a general-purpose or specialpurpose computing system or combinations. While not all computing systems require a user interface, in some embodiments, the computing system includes a user interface system for use in interfacing with a user. User interfaces act as input or output mechanism to users for instance via displays.

[0117] Those skilled in the art will appreciate that at least parts of the invention may be practiced in network computing environments with many types of computing system configurations, including, personal computers, desktop computers, laptop computers, message processors, hand-held devices, multi-processor systems, microprocessor-based or programmable consumer electronics, network PCs, minicomputers, mainframe computers, mobile telephones, PDAs, pagers, routers, switches, datacenters, wearables, such as glasses, and the like. The invention may also be practiced in distributed system environments where local and remote computing system, which are linked, for example, either by hardwired data links, wireless data links, or by a combination of hardwired and wireless data links, through a network, both perform tasks. In a distributed system environment, program modules may be located in both local and remote memory storage devices.

[0118] Those skilled in the art will also appreciate that at least parts of the invention may be practiced in a cloud computing environment. Cloud computing environments may be distributed, although this is not required. When distributed, cloud computing environments may be distributed internationally within an organization and / or have components possessed across multiple organizations. In this description and the following claims, “cloud computing” is defined as a model for enabling on-demand network access to a shared pool of configurable computing resources, e.g., networks, servers, storage, applications, and services. The definition of “cloud computing” is not limited to any of the other numerous advantages that can be obtained from such a model when deployed. The computing systems of the figures include various components or functional blocks that may implement the various embodiments disclosed herein as explained. The various components or functional blocks may be implemented on a local computing system or may be implemented on a distributed computing system that includes elements resident in the cloud or that implement aspects of cloud computing. The various components or functional blocks may be implemented as software, hardware, or a combination of software and hardware. The computing systems shown in the figures may include more or less than the components illustrated in the figures and some of the components may be combined as circumstances warrant.

[0119] Any reference signs in the claims should not be construed as limiting the scope. | BASF SE | 230213 | 230213WQ01 ~

[0120] The invention refers to a controlling and / or monitoring of an asset transfer. Time series monitoring data related to the asset is provided. A trained data-driven model is provided. The data-driven model partially includes discrete weights, wherein the model has been trained i) by providing a training data set including a) control and / or monitoring data of asset transfers and b) time series monitoring data related to the asset, ii) by initializing the model on a classical computer based on the training data set and iii) by providing the initialized model for determining the discrete weights based on the training data set to a quantum computer. The trained model determines control and / or monitoring data for controlling and / or monitoring an asset transfer based on the time series monitoring data. Control and / or monitoring data are generated based on the trained data-driven model and the time series monitoring data related to the asset.

Claims

BASF SE230213Claims:1 . A computer-implemented method for controlling and / or monitoring one or more asset transfer(s), wherein the method comprises:- provide time series monitoring data related to the asset,- provide trained data-driven model, wherein the data-driven model at least partially includes discrete weights, wherein the data-driven model has been trained i) by providing a training data set including a) control and / or monitoring data of asset transfers and b) time series monitoring data related to the asset, ii) by initializing the model on a classical computer based on the training data set and iii) by providing the initialized model for determining the discrete weights based on the training data set to a quantum computing unit, wherein the trained model determines a model output indicative of control and / or monitoring data for controlling and / or monitoring an asset transfer based on the time series monitoring data,- generate control and / or monitoring data based on the trained data-driven model and the time series monitoring data related to the asset, and- provide control and / or monitoring data for controlling and / or monitoring a transfer of the asset.

2. The method according to claim 1 , wherein the asset includes a digitalized representation of a product asset associated with production and / or product performance of a product.

3. The method according to any of the preceding claims, wherein the time series monitoring data related to the asset is associated with a transfer value of the asset transfer.

4. The method according to any of the preceding claims, wherein the data-driven model is a boosting classifier comprising a plurality of weak classifiers and wherein the discreet weights determine respective contributions of classifications of the weak classifiers to the classification of the boosting classifier.

5. The method according to any of the preceding claims, wherein the training of the data-driven model on the quantum computer comprises removing correlations in the initialized model above a predetermined threshold.BASF SE2302136. The method according to any of the preceding claims, wherein the method further comprises aggregating the received time series monitoring data for a predetermined time interval into a feature set representing the time series monitoring data for that predetermine time interval and wherein the data-driven model is trained to generate the model output based on the feature set for the predetermined time interval as input.

7. The method according to claim 6, wherein the feature set of a predetermined time interval comprises as feature at least one of: a simple moving average, a relative strength index, a rate of change, and a moving average converge divergence.

8. The method according to any of the preceding claims, wherein the data-driven model comprises an asset value part comprising at least partly the discreet weights correlated to providing a binary output associated with an asset value, and a control data part for generating the model output based on the binary output of the asset value part.

9. The method according to any of the preceding claims, wherein the method further comprises generating a plurality of data-driven models, wherein each data-driven model comprises a different set of model parameters defining the model, wherein the method further comprises selecting from the plurality of machine learning based models a machine learning based model based on a validation of the data-driven models.

10. A computer-implemented method for generating control data for controlling a quantum computer to generate a data-driven model, wherein the data-driven model is configured to generate control and / or monitoring data suitable to control and / or monitor an asset transfer based on time series monitoring data related to the asset, wherein the method comprises:- provide a training data set comprising a) control and / or monitoring data of asset transfers and b) time series monitoring data related to the asset,- initialize a data-driven model at least partially including discrete weights, wherein the model is initialized based on the training data set, and- generate control data to control a quantum computer for training the data-driven model by determining the discreet weights of the data-driven model based on the initialized model and the training data set, wherein the trained model determines a model output indicativeBASF SE230213of control and / or monitoring data for controlling and / or monitoring an asset transfer based on the time series monitoring data.

11. An apparatus for controlling and / or monitoring one or more asset transfer(s), wherein the apparatus comprises one or more processors configured to:- provide time series monitoring data related to the asset,- provide trained data-driven model, wherein the data-driven model at least partially includes discrete weights, wherein the data-driven model has been trained i) by providing a training data set including a) control and / or monitoring data of asset transfers and b) time series monitoring data related to the asset, ii) by initializing the model on a classical computer based on the training data set and iii) by providing the initialized model for determining the discrete weights based on the training data set to a quantum computing unit, wherein the trained model determines a model output indicative of control and / or monitoring data for controlling and / or monitoring an asset transfer based on the time series monitoring data,- generate control and / or monitoring data based on the trained data-driven model and the time series monitoring data related to the asset, and- provide control and / or monitoring data for controlling and / or monitoring a transfer of the asset.

12. An apparatus for generating control data for controlling a quantum computer to generate a data-driven model, wherein the data-driven model is configured to generate control and / or monitoring data suitable to control and / or monitor an asset transfer based on time series monitoring data related to the asset, wherein the apparatus comprises one or more processors configured to:- provide a training data set comprising a) control and / or monitoring data of asset transfers and b) time series monitoring data related to the asset,- initialize a data-driven model at least partially including discrete weights, wherein the model is initialized based on the training data set, and- generate control data to control a quantum computer for training the data-driven model by determining the discreet weights of the data-driven model based on the initialized model| BASF SE | 230213 | 230213WQ01 ~ and the training data set, wherein the trained model determines a model output indicative of control and / or monitoring data for controlling and / or monitoring an asset transfer based on the time series monitoring data.

13. A system for controlling and / or monitoring one or more asset transfer(s), wherein the system comprises: a quantum computer unit, an apparatus according to claim 12 for generating control data for controlling the quantum computer to train a data-driven model, and an apparatus according to claim 11 utilizing the data-driven model for providing con- trol and / or monitoring data for controlling and / or monitoring a transfer of the asset.

14. A computer program product for controlling and / or monitoring one or more asset transfer(s) comprising program code means for causing an apparatus according to claim 11 to carry out a method according to any of claims 1 to 9.

15. A computer program product for generating control data for controlling a quantum computer to generate a data-driven model comprising program code means for causing an apparatus according to claim 12 to carry out a method according to claims 10.

Images