Model learning device, control device, model learning method, and computer program
The model learning device stabilizes system control by using bijective mappings in a nonlinear state equation, ensuring output components become 0 when input components are 0, thus preventing system instability and enabling accurate future state prediction.
Patent Information
- Application Number
- JP2023076233
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Filing Date
- 2023-05-02
- Publication Date
- 2024-11-15
- Estimated Expiration
- 2043-05-02
AI Technical Summary
Existing model learning devices face instability when controlling systems due to the lack of guarantee that the output component becomes 0 when the input component is 0, leading to potential system instability.
A model learning device that incorporates bijective mappings Ψ and Φ within a nonlinear state equation, ensuring that when at least one component of the input or output variable is 0, the corresponding component in the other variable also becomes 0, using multilayer neural networks to stabilize system control.
The solution ensures stable system control by guaranteeing that the output component becomes 0 when the input component is 0, allowing for accurate prediction of future system states and optimal control.
Smart Images

Figure 2024160826000001_ABST
Abstract
Description
[Technical field]
[0001] The present invention relates to a model learning device, a control device, a model learning method, and a computer program. [Background technology]
[0002] Conventionally, model learning devices are known that learn a model that represents the relationship between an input for controlling a system and an output from the system in response to the input. For example, Patent Document 1 discloses a technique for learning a model that includes a bijective mapping that transforms an input so that the output falls within a predetermined region. Patent Document 2 discloses a model learning device that learns a model that includes two bijective mappings that transform input and output and a linear dynamic equation. [Prior art documents] [Patent documents]
[0003] [Patent Document 1] Patent Publication No. 2022-016709 [Patent Document 2] Patent Publication No. 2021-189934 Summary of the Invention [Problem to be solved by the invention]
[0004] However, even with the above-mentioned prior art, there is still room for improvement in the technology of learning a model capable of stably controlling a system in a model learning device. For example, neither Patent Document 1 nor Patent Document 2 describes that a bijective mapping included in a model learned by a model learning device has a value of 0 as a corresponding output component when an input component is 0. Therefore, in the model learned by the model learning device described in Patent Document 1 and the model learned by the model learning device described in Patent Document 2, there is a risk that the corresponding output component will not become 0 even if the input component becomes 0. If such a model is used to control a system, there is a risk that the system will become unstable.
[0005] The present invention has been made to solve the above-mentioned problems, and aims to provide a technology that can learn a model that can stably control a system in a model learning device that learns a model that represents the relationship between input and output in a system. [Means for solving the problem]
[0006] The present invention has been made to solve the above-mentioned problems, and can be realized in the following forms.
[0007] (1) According to one aspect of the present invention, there is provided a model learning device that learns a model that represents a relationship between an input variable v (vector) input to a system and an output variable y (vector) output from the system. The model learning device includes a model storage unit that stores a model used in learning a nonlinear state equation for predicting the output variable y using the input variable v, and a learning unit that learns the state equation using a model stored in the model storage unit and an input / output data set that includes a plurality of pairs of input variable data and output variable data for the model, the model being a state equation that includes a bijective mapping Ψ with the input variable v as an input and a bijective mapping Φ with the output variable y as an input, the mapping Ψ having at least one output component that is 0 when at least one component of the input variable v is 0, and the mapping Φ having at least one output component that is 0 when at least one component of the output variable y is 0.
[0008] According to this configuration, the model is a state equation including a bijective mapping Ψ with an input variable v input to the system as an input, and a bijective mapping Φ with an output variable y output from the system as an input. Such a state equation can be linearized by making each of the mappings Ψ and Φ an internal variable, so that even in a model with a nonlinear structure, it is possible to guarantee that the solution of an optimal control problem for finding an optimal input that achieves an output target is unique. In addition, when at least one component of the input variable v is 0, at least one output component of the mapping Ψ becomes 0, and when at least one component of the output variable y is 0, at least one output component of the mapping Φ becomes 0. As a result, it is guaranteed that when a specific input component becomes 0, the corresponding output component of the state equation including each of the mappings Ψ and Φ becomes 0. Therefore, the model learning device can learn a model that can stably control the system.
[0009] (2) In the model learning device of the above form, the model may be defined by equation (1), the mapping Ψ may be defined by equations (2) to (4) representing a multi-layer neural network, and the mapping Φ may be defined by equations (5) to (7) representing a multi-layer neural network.
number
number
number
number
number
number
number
[0010] (3) In the model learning device of the above form, the model may be defined by equation (1), the mapping Ψ may be defined by equations (8) to (11) representing a multilayer neural network whose intermediate layer is represented by at least one of equations (9) and (10), and the mapping Φ may be defined by equations (12) to (15) representing a multilayer neural network whose intermediate layer is represented by at least one of equations (13) and (14).
number
number
number
number
number
number
number
number
number
[0011] (4) In the model learning device of the above form, the learning unit may learn the state equation defined by equation (18) by defining the mapping Ψ by equation (16) using an internal variable u in equation (1) and defining the mapping Φ by equation (17) using an internal variable x.
number
number
number
[0012] (5) In the model learning device of the above form, the learning unit may learn a state equation shown in equation (21) obtained by discretizing equation (18) at a time step of discrete time k, by discretizing equation (16) at a time step of discrete time k into equation (19), and by discretizing equation (17) at a time step of discrete time k into equation (20).
number
number
number
[0013] (6) In the model learning device of the above embodiment, the learning unit may provide the model with the set of input variable data from the input / output data set to estimate an output, evaluate the degree of agreement between the estimated output and the set of output variable data from the input / output data set, and update learning parameters of the model according to the evaluation result, thereby learning the state equation. According to this configuration, the learning unit evaluates the degree of agreement between the output estimated using the input variable data set from the input / output data set and the output variable data set. The learning unit updates learning parameters for the model according to the evaluation of the degree of agreement, and learns the state equation. That is, the learning unit can learn a nonlinear state equation according to a learning method using a previously prepared input / output data set as teacher data. This makes it possible to learn a model that is in line with the actual system, thereby improving the tracking ability of the output from the system to the target value. Therefore, the model learning device can learn a model that can stably control the system and has high tracking ability of the output from the system to the target value.
[0014] (7) According to another aspect of the present invention, there is provided a control device for controlling a system. This control device includes the above-mentioned model learning device, a determination unit that determines a target value of the input variable v corresponding to a target value of the output variable y using the state equation learned by the learning unit, and a control unit that controls the system by inputting the target value of the input variable v determined by the determination unit. According to this configuration, the control unit controls the system by inputting the target value of the input variable v determined using a model that guarantees that when a specific input becomes 0, the corresponding output becomes 0. This allows the control device to stably control the system.
[0015] (8) According to yet another aspect of the present invention, a model expressing a relationship between an input variable v (vector) input to a system and an output variable y (vector) output from the system is learned. The model learning method includes an acquisition step of acquiring a model used in learning a nonlinear state equation for predicting the output variable y using the input variable v, and a learning step of learning the state equation using the model acquired in the acquisition step and an input / output data set including a plurality of pairs of input variable data and output variable data for the model, the model being a state equation including a bijective mapping Ψ having the input variable v as an input and a bijective mapping Φ having the output variable y as an input, the mapping Ψ having at least one output component that is 0 when at least one component of the input variable v is 0, and the mapping Φ having at least one output component that is 0 when at least one component of the output variable y is 0. According to this configuration, the trained model is a state equation including a bijective mapping Ψ with an input variable v input to the system as an input, and a bijective mapping Φ with an output variable y output from the system as an input, so that it is guaranteed that the solution to the optimal control problem for finding the optimal input that realizes the output target is unique. In addition, in both mappings Ψ and Φ, it is guaranteed that when a specific input becomes 0, the corresponding output becomes 0. As a result, by using the model trained in the training process, when an input component becomes 0, the corresponding output component becomes 0. Therefore, it is possible to train a model that can stably control the system.
[0016] (9) According to yet another aspect of the present invention, there is provided a computer program for causing an information processing device to execute learning of a model representing a relationship between an input variable v (vector) input to a system and an output variable y (vector) output from the system. This computer program causes the information processing device to execute an acquisition function for acquiring a model used in learning a nonlinear state equation for predicting the output variable y using the input variable v, and a learning function for learning the state equation using the model acquired by the acquisition function and an input / output data set including a plurality of pairs of input variable data and output variable data for the model, the model being a state equation including a bijective mapping Ψ having the input variable v as an input, and a bijective mapping Φ having the output variable y as an input, the mapping Ψ having at least one output component that is 0 when at least one component of the input variable v is 0, and the mapping Φ having at least one output component that is 0 when at least one component of the output variable y is 0. According to this configuration, the model learned by the information processing device is a state equation including a bijective mapping Ψ with an input variable v input to the system as an input, and a bijective mapping Φ with an output variable y output from the system as an input, so that it is guaranteed that the solution of the optimal control problem for finding the optimal input that realizes the output target is unique. In addition, in both mappings Ψ and Φ, it is guaranteed that when a specific input becomes 0, the corresponding output becomes 0. As a result, by using a model learned by a learning function, when an input component becomes 0, the corresponding output component becomes 0. Therefore, it is possible to learn a model that can stably control the system.
[0017] The present invention can be realized in various forms, such as an apparatus and method for learning a model of a nonlinear system, an apparatus and method for estimating a state using a model obtained by learning, a system including these apparatuses, a computer program executed in these apparatuses and systems, a server device for distributing the computer program, a non-transitory storage medium on which the computer program is stored, and the like. [Brief description of the drawings]
[0018] [Figure 1] FIG. 1 is a schematic diagram showing a configuration of a model learning device according to a first embodiment. [Diagram 2] 4 is a flowchart of a model learning method according to the first embodiment. [Diagram 3] FIG. 13 is a diagram showing a change in an input value over time. [Figure 4] FIG. 11 is a diagram comparing predicted values using the model of the first embodiment with actual measured values. [Diagram 5] FIG. 11 is a diagram comparing predicted values and actual measured values using a model of a comparative example. [Figure 6] FIG. 13 is a schematic diagram showing the configuration of a model learning device according to a second embodiment. [Figure 7] FIG. 13 is a schematic diagram showing the configuration of a model learning device according to a third embodiment. [Figure 8] FIG. 13 is a schematic diagram showing the configuration of a control device according to a fourth embodiment. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0019] First Embodiment FIG. 1 is a schematic diagram showing the configuration of a model learning device 100 of the first embodiment. The model learning device 100 of the present embodiment is a device that learns a model of a nonlinear system. Here, the "nonlinear system" means a system having a property that the relationship between input and output parameters for an arbitrary controlled object (system) cannot be expressed by a linear equation or cannot be approximated. In addition, in this embodiment, a nonlinear state equation is exemplified as the "model". That is, the model learning device 100 learns a nonlinear state equation that predicts the output variable y of the system as a result of control by an input variable v input to the system by regarding the state of the arbitrary system as an output variable y output from the system. Note that the "state equation" means an equation that determines its own output variable y·(t) by the output variable y(t) at the current time t, such as "y·(t)=f(y(t),···)". Hereinafter, for convenience of notation, the time derivative of an arbitrary variable z is described as "z·".
[0020] The system includes, for example, a driving engine such as an internal combustion engine, a hybrid engine, a powertrain, a fuel cell, an induction motor, etc. When a driving engine such as an internal combustion engine is considered as a system, the model learned by the model learning device 100 is, for example, a nonlinear state equation that expresses the relationship between the operation amount of the actuator of the controlled part, the disturbance to the controlled part, the state of the controlled part, the output of the controlled part, the output target value of the controlled part, etc. More specifically, when an internal combustion engine mounted on a vehicle is assumed as a system, the model learning device 100 receives input variables v such as the accelerator opening and the EGR valve opening, and learns state equations that predict, as output variables y, the output value of the internal combustion engine, the amount of carbon dioxide emission, and the amount of hydrocarbon emission output from the internal combustion engine.
[0021] When a hybrid engine of an internal combustion engine and a motor mounted on a vehicle is assumed as a system, model learning device 100 learns a state equation that predicts, as output variable y, the output value of the internal combustion engine, the output value of the motor, the amount of stored electricity in the battery, and the limit value of the amount of stored electricity, which are output from the hybrid engine, by inputting, as input variable v, the opening degree of the accelerator, the amount of brake operation, the acceleration of the vehicle, etc. When an internal combustion engine or a hybrid engine mounted on a vehicle is assumed as a system, the initial condition is the running state of the vehicle that changes as the vehicle runs (whether it is turning, whether it is going up a slope, etc.).
[0022] When a fuel cell is assumed as the system, the model learning device 100 receives an amount of fuel supply as an input variable v and learns a state equation that predicts generated power as an output variable y. When an induction motor is assumed as the system, the model learning device 100 receives a power supply voltage as an input variable v and learns a state equation that predicts the rotation speed of the motor as an output variable y. Note that the field to which the model learning device 100 of the present embodiment is applied is not limited to this.
[0023] The model learning device 100 is implemented, for example, as a personal computer (PC). The model learning device 100 is a multi-function peripheral device (PC) that includes a CPU 110, a storage unit 120, a ROM / RAM 130, a communication unit 140, and an input / output unit 150. The individual units of the model learning device 100 are connected to each other via a bus.
[0024] The CPU 110 includes a control unit 111 and a learning unit 112. The control unit 111 loads a computer program stored in the ROM 130 into the RAM 130 and executes it. By this, the learning unit 112 controls each unit of the model learning device 100. The learning unit 112 learns a nonlinear state equation for predicting an output variable y that represents the state of an arbitrary system (nonlinear system). The function of the learning unit 112 will be described in detail later.
[0025] The storage unit 120 is a storage medium constituted by a hard disk, a flash memory, a memory card, etc. The storage unit 120 has a model storage unit 121 and a dataset storage unit 122.
[0026] The model storage unit 121 stores in advance a model used for learning the state equation by the learning unit 112. In this embodiment, the model stored in the model storage unit 121 is a state equation including a bijective mapping Ψ with an input variable v as an input, and a bijective mapping Φ with an output variable y as an input, and is defined by formula (1). Here, "bijection" means that when the result of mapping a set A is set B, each element of A and B necessarily have a one-to-one correspondence. This means that, for example, when a function f is bijective, there is a unique inverse function f -1 is synonymous with the existence of
number
[0027] In this embodiment, the bijective mapping Ψ included in the formula (1) is defined by the following formulas (2) to (4) which represent a multilayer neural network.
number
number
number
[0028] In this embodiment, the bijective mapping Φ included in formula (1) is defined by formulas (5) to (7) which represent a multi-layer neural network.
number
number
number
[0029] The data set storage unit 122 pre-stores an input / output data set including a plurality of pairs of input variable data and output variable data for the model of formula (1). The pairs of input variable data and output variable data are obtained in advance by experiments and calculations on the system. The input / output data set is used as teacher data used for learning the state equation by the learning unit 112. Hereinafter, among the input / output data sets, a plurality of input variable data will be collectively referred to as the "input variable data set", and a plurality of output variable data will be collectively referred to as the "output variable data set".
[0030] The communication unit 140 controls communication between the model learning device 100 and other devices via a communication interface. Examples of the other devices include a control device that controls the system, another information processing device, and a measuring instrument for acquiring an input / output data set from the data set storage unit 122. The input / output unit 150 is various interfaces used for inputting and outputting information between the model learning device 100 and a user. Examples of the input / output unit 150 include a touch panel, a keyboard, a mouse, an operation button, and a microphone as an input unit, and a touch panel, a monitor, a speaker, and an LED (Light Emitting Diode) indicator as an output unit.
[0031] FIG. 2 is a flowchart of the model learning method of the first embodiment. The model learning method in the model learning device 100 is executed, for example, in response to a request from a user, such as the start of a predetermined application. In this embodiment, in the state equation shown in equation (1), a function form of the function F shown in equation (22) is learned (estimated) using a known input / output data set including the output variable y, the input variable v, the exogenous input d in the system, the time derivative y· of the output variable y, and the time derivative d· of the exogenous input d. Here, the output variable y is an n-dimensional vector. where the input variables v are n-dimensional vectors and the exogenous input d is a p-dimensional vector.
number
[0032] First, the learning unit 112 acquires a model stored in the model storage unit 121 (step S11). Specifically, the learning unit 112 assumes that the model for learning the function F is the state equation shown in formula (1). The learning unit 112 initializes each variable in the state equation shown in formula (1) by setting the value of each variable to zero or a random value.
number
[0033] In this embodiment, the learning unit 112 defines the mapping Ψ included in formula (1) as the internal variable u shown in formula (16), and defines the mapping Φ included in formula (1) as the internal variable x shown in formula (17). As a result, the learning unit 112 learns the state equation of formula (18) in which formula (1) is represented by the internal variables u and x. The effect of defining the mappings Φ and Ψ included in the state equation of formula (1) by the internal variables x and u, respectively, will be described later.
number
number
number
number
number
[0034] In this embodiment, the learning unit 112 reads state equations shown in equations (19) to (21) obtained by discretizing equations (16) to (18) at a predetermined time step Δt at a discrete time k.
number
number
number
[0035] Next, the learning unit 112 acquires an input / output data set [y, v, d, y·, d·] for the output variable y, the input variable v, the exogenous input d, the time differential y· of the output variable y, and the time differential d· of the exogenous input d from the data set storage unit 122 (step S12). In this embodiment, j sets (j is a natural number, j=1 to N) of data are included in the input / output data set [y, v, d, y·, d·]. Of the acquired input / output data set, [y j , v j , d j , d· j ] corresponds to the input variable data set, and [y· j ] corresponds to the output variable data set.
[0036] Next, the learning unit 112 provides the model with an input data set and estimates the output (step S13). Specifically, the learning unit 112 estimates the output by applying the input variable data set [y j , v j , d j , d· j ]. This allows us to obtain the estimate of the output variable y·j (the left side of equation (25)). Note that (∂Φ / ∂y) -1 is a function of the output variable y and the exogenous input d, so the output variable y j and exogenous input d jIn addition, (∂Φ / ∂d) on the right-hand side of equation (17) is a function of the input variable v and the exogenous input d, so the input variable v j and exogenous input d j It can be evaluated by substituting
number
[0037] Next, the learning unit 112 evaluates the degree of agreement between the estimated output and the output variable data set (step S14). j The estimated values of and the output variable data set obtained in step S12 [y· j ] is evaluated. The learning unit 112 can use, for example, the mean square error (MSE) shown in equation (26) as an index of the degree of match. In the case of MSE, the smaller the value of J on the left side of the equal sign, the higher the degree of match. Note that the learning unit 112 may evaluate the degree of match using an index such as the absolute mean error rate or cross entropy instead of the mean square error.
number
[0038] Next, the learning unit 112 determines whether the degree of agreement is sufficient (step S15). For example, when the MSE of equation (18) is used, the learning unit 112 can determine that the degree of agreement is sufficient when the value of J is equal to or less than a predetermined value. The learning unit 112 may also determine that the degree of agreement is sufficient when the rate of change of the value of J is equal to or less than a predetermined value. The predetermined value can be determined arbitrarily.
[0039] If the degree of agreement is not sufficient (step S15: NO), the learning unit 112 proceeds to step S16, and determines, for example, the function A included in the equation (1), the weights W included in the equations (2) to (7) in the model of the equation (1) defined in step S11. Ψ , W Φ or bias b Ψ , bΦ The learning unit 112 may update the learning parameters such as J, J, and the like. For example, the learning unit 112 may evaluate the gradient of J for each learning parameter by backpropagation, and update each learning parameter based on various gradient methods. After that, the learning unit 112 proceeds to step S13, and repeats the estimation and evaluation of the output.
[0040] If the degree of agreement is sufficient (step S15: YES), the learning unit 112 ends the process. At this time, the learning unit 112 may output the learned function F to the input / output unit 150. It may also be stored in the storage unit 120 or transmitted to another device via the communication unit 140.
[0041] When model learning device 100 of this embodiment is combined with a control device that controls the manipulated variable of a system, model learning device 100 outputs function F learned in learning unit 112 to the control device. The control device uses the output function F to calculate an input for controlling a future output from the output of the system at the current time. The control device outputs the calculated input to the system and controls the system.
[0042] In the model learned by the model learning device 100 of this embodiment, it is guaranteed that when a specific input component becomes 0, the corresponding output component becomes 0. This makes it possible to stably control the system to be controlled. Here, the results of studying a model related to the intake and exhaust system of an internal combustion engine as an example of a model learned by the model learning device 100 of this embodiment will be described. In this study, the items shown in equation (27) were used as the input variable v, and the items shown in equation (28) were used as the output variable y.
number
number
[0043] Fig. 3 is a diagram showing changes over time in input variables. The relationship between the EGR valve opening (input variable) and the EGR rate (output variable) in the model learned by the model learning device 100 will be specifically described. Fig. 3 shows an example of changes over time in the EGR valve opening in an internal combustion engine. In the changes over time in the EGR valve opening shown in Fig. 3, the EGR valve opening is 0 between time t1 and time t2.
[0044] FIG. 4 is a diagram comparing a predicted value using a model learned by the model learning device 100 of the present embodiment with an actual measured value. In FIG. 4, the time change of the EGR rate predicted using the model learned by the model learning device 100 with respect to the time change of the EGR valve opening shown in FIG. 3 is indicated by a dashed line Pv1, and the time change of the actual EGR rate is indicated by a solid line Av1. As shown in FIG. 4, during the period from time t1 to time t2 when the EGR valve opening shown in FIG. 3 becomes 0, the EGR rate predicted using the model is also 0. In this way, in the model learned by the model learning device 100, when a specific input variable v (v1: EGR valve opening) becomes 0, the corresponding output variable y (y1: EGR rate) becomes 0.
[0045] Fig. 5 is a diagram comparing predicted values and actual measured values using a model learned by the model learning device of the comparative example. In Fig. 5, the time change of the EGR rate predicted using the model learned by the model learning device of the comparative example with respect to the time change of the EGR valve opening shown in Fig. 3 is shown by a dashed dotted line Pv0, and the time change of the actual EGR rate is shown by a solid line Av1. The model learned by the model learning device of the comparative example is the state equation shown in formula (1), but the mappings Ψ and Φ are not defined as in formulas (2) to (4) and formulas (5) to (7).
[0046] As shown in Fig. 5, the EGR rate predicted using the model learned by the model learning device of the comparative example is not 0 during the period from time t1 to time t2 when the EGR valve opening shown in Fig. 3 becomes 0. In this way, with the model learned by the model learning device of the comparative example, there is no guarantee that the output variable y will become 0 when the input variable v becomes 0. For this reason, when the system is controlled using such a model, even if a specific input variable v (v1: EGR valve opening) becomes 0, the corresponding output variable y (y1: EGR rate) does not become 0, which may cause the system control to become unstable.
[0047] According to the model learning device 100 of the present embodiment described above, the model learned by the model learning device 100 is a state equation including a bijective mapping Ψ with an input variable v input to the system as an input, and a bijective mapping Φ with an input variable y output from the system as an input. Such a state equation can be linearized by making each of the mappings Ψ and Φ an internal variable, so that even in a model with a nonlinear structure, it is possible to guarantee that the solution of an optimal control problem for finding an optimal input that achieves a target value of the output is unique. In addition, when at least one component of the input variable v is 0, at least one output component of the mapping Ψ becomes 0, and when at least one component of the output variable y is 0, at least one output component of the mapping Φ becomes 0. This ensures that when a specific input component becomes 0, the corresponding output component becomes 0, making it difficult for unintended control to occur in the control of the system using the model. Therefore, the model learning device 100 can learn a model that can stably control the system. In other words, the model learning device 100 can learn a model that can construct a control device that stably controls a system.
[0048] Moreover, according to the model learning device 100 of this embodiment, each of the mappings Ψ and Φ included in the model of equation (1) includes an exogenous input d, which is an uncontrollable input that affects the change in the output variable y. As a result, the model of equation (1) becomes a state equation that also takes into account the influence of such an exogenous input d, and by using this model, it is possible to predict the future state of the system with high accuracy. Moreover, the mapping Ψ represented by the multilayer neural network defined by equations (2) to (4) is expressed by the input variable v in equation (3) that represents the intermediate layer of the multilayer neural network. (i-1) When becomes 0, the input variable v (i) In other words, the output of the mapping Ψ is also 0. In addition, the mapping Φ represented by the multilayer neural network defined by equations (5) to (7) is expressed by the output variable y (i-1) When becomes 0, the output variable y (i) also becomes 0. That is, the output of the mapping Φ also becomes 0. In this way, since the model expressed by equation (1) includes the mapping Ψ expressed by equations (2) to (4) and the mapping Φ expressed by equations (5) to (7), it is guaranteed that when a specific input becomes 0, the corresponding output becomes 0. Therefore, model learning device 100 can learn a model that can stably control the system.
[0049] According to the model learning device 100 of this embodiment, the mapping Ψ expressed by the multilayer neural network defined by equations (2) to (4) is such that, in equation (3) showing the intermediate layer, if all elements of the input variable v (vector) are 0, then equation (2) becomes 0. Ψ is a diagonal matrix, if the j-th element of the input variable v is 0, the j-th element of the mapping Ψ is also 0. In addition, in the mapping Φ expressed by the multilayer neural network defined by equations (5) to (7), if all elements of the output variable y (vector) are 0, then equation (6) representing the intermediate layer becomes 0. Weight W Φis a diagonal matrix, if the j-th element of the output variable y is 0, then the j-th element of the mapping Φ is also 0. Due to these properties, the model expressed by equation (1) has origin invariance. Here, "origin invariance" refers to the property that when the function A included in equation (1) is a stable matrix, that is, when the real parts of all eigenvalues are negative, if all elements of the input variable v are kept constant at 0, then the output variable y will always converge to 0 for all elements. More specifically, if the input variable v is 0, then the internal variable u included in equation (16) is 0, and the time derivative x· of the internal variable x included in equation (17) is 0 when the internal variable x is 0, so the output variable y is 0. In this way, by using the model learned in the learning unit 112, when the input variable v becomes 0, the corresponding output variable y becomes 0. Therefore, the model learning device 100 can learn a model that can stably control the system.
[0050] Moreover, according to the model learning device 100 of this embodiment, the learning unit 112 defines the mapping Ψ in equation (1) by equation (16) using the internal variable u, and defines the mapping Φ by equation (17) using the internal variable x in equation (1), thereby learning the state equation defined by equation (18). As a result, as shown in equation (18), the state equation shown in equation (1) can be linearized, and therefore it is possible to guarantee that the state equation shown in equation (1) provides a unique solution to the optimal control problem for obtaining an optimal input that achieves the target value of the output. Therefore, the model learning device 100 can learn a model that can stably control the system.
[0051] Moreover, according to the model learning device 100 of this embodiment, the learning unit 112 learns the state equation shown in equation (21) obtained by discretizing equation (18) at the time step of discrete time k by discretizing equation (16) at the time step of discrete time k as equation (19) and equation (17) at the time step of discrete time k as equation (20). This makes it possible to shorten the time required for model learning by making the number of internal variables x and u finite. Therefore, the model learning device 100 can learn a model capable of constructing a control device that stably controls a system in a relatively short time.
[0052] Moreover, according to the model learning device 100 of the present embodiment, the learning unit 112 evaluates the degree of agreement between the output estimated using the input variable data set among the input / output data sets and the output variable data set. The learning unit 112 updates the learning parameters for the model according to the evaluation of the degree of agreement, and learns the state equation. That is, the learning unit 112 can learn the nonlinear state equation according to a learning method using the input / output data set prepared in advance as teacher data. This makes it possible to learn a model that is in line with the actual system. Therefore, the model learning device 100 can learn a model that can stably control the system and has high tracking ability for the target value of the output from the system.
[0053] Moreover, according to the model learning method of the present embodiment, the learned model is a state equation including a bijective mapping Ψ with an input variable v input to the system as an input, and a bijective mapping Φ with an output variable y output from the system as an input, so that it is possible to guarantee that the solution of the optimal control problem for finding an optimal input that achieves a target value of the output is unique. In addition, in both mappings Ψ and Φ, when a specific input becomes 0, the corresponding output becomes 0. As a result, by using the model learned by the model learning method shown in FIG. 2, when an input component becomes 0, the corresponding output component becomes 0. Therefore, the model learning device 100 can learn a model that can stably control the system.
[0054] Furthermore, according to the computer program of this embodiment, the derived model is a state equation including a bijective mapping Ψ with an input variable v input to the system as an input, and a bijective mapping Φ with an output variable y output from the system as an input, so that it is possible to guarantee that the solution is unique. In addition, in both mappings Ψ and Φ, when a specific input becomes 0, the corresponding output becomes 0. As a result, by using the model learned by the CPU 110, when an input component becomes 0, the corresponding output component becomes 0. Therefore, the model learning device 100 can learn a model that can stably control the system.
[0055] <Second embodiment> 6 is a schematic diagram showing the configuration of a model learning device 200 according to the second embodiment. Compared with the model learning device 100 (FIG. 1) according to the first embodiment, the model learning device 200 according to the second embodiment differs in the configuration of the model stored in the model storage unit.
[0056] The model learning device 200 is, for example, a personal computer (PC). The model learning device 200 is a general-purpose computing device (CPU) 110, a storage unit 220, a ROM / RAM 130, a communication unit 140, and an input / output unit 150. The components of the model learning device 200 are connected to each other via a bus. The storage unit 220 is a storage medium constituted by a hard disk, a flash memory, a memory card, or the like. The storage unit 220 has a model storage unit 221 and a dataset storage unit 122.
[0057] The model storage unit 221 prestores a model used for learning of a state equation by the learning unit 112. In this embodiment, the model stored in the model storage unit 221 is a state equation including a bijective mapping Ψ with an input variable v as an input, and a bijective mapping Φ with an input variable y as an input, and is defined by Equation (1).
number
[0058] In this embodiment, the bijective mapping Ψ included in formula (1) is defined by formulas (8) to (11) that represent a multilayer neural network whose intermediate layer is represented by at least one of formulas (9) and (10).
number
number
number
number
[0059] In this embodiment, the bijective mapping Φ included in formula (1) is defined by formulas (12) to (15) which represent a multilayer neural network whose intermediate layer is represented by at least one of formulas (13) and (14).
number
number
number
number
[0060] The model learning method in the model learning device 200 is performed by the same method as the model learning method of the first embodiment. Specifically, first, the learning unit 112 acquires a model stored in the model storage unit 221. In this embodiment, the learning unit 112 acquires a model of formula (1) including a mapping Ψ defined by formulas (8) to (11) and a mapping Φ defined by formulas (12) to (15). After that, the learning unit 112 acquires an input / output data set, and gives an input data set to the model to estimate an output. The learning unit 112 evaluates the degree of agreement between the estimated output and the output variable data set. If the degree of agreement is not sufficient, the learning unit 112 repeats the estimation and evaluation of the output. If the degree of agreement is sufficient, the learning unit 112 ends the process.
[0061] According to the model learning device 200 of this embodiment described above, each of the mappings Ψ and Φ included in the model of equation (1) learned by the model learning device 200 includes an exogenous input d, which is an uncontrollable input that affects changes in the output variable y, and therefore, by using the model of equation (1), it is possible to predict the future state of the system with high accuracy. Moreover, the mapping Ψ represented by the multilayer neural network defined by equations (8) to (11) has an intermediate layer represented by at least one of equations (9) and (10). In equation (9), n, which is a part of the input variable v, a dimensional input variable v a (i-1) When becomes 0, the input variable v a (i) Also, in equation (10), n a dimensional input variable v a (i-1) , or n, which is a part of the input variables v b dimensional input variable v b (i-1) When becomes 0, the input variable v a (i) Therefore, part of the output of the mapping Ψ is also 0. In addition, the mapping Φ represented by the multilayer neural network defined by equations (12) to (15) has an intermediate layer represented by at least one of equations (13) and (14). In equation (13), n c dimensional output variable y c (i-1) When becomes 0, the output variable y c (i) Also, in equation (14), n c dimensional output variable y c (i-1) , or n which is part of the output variable y d dimensional output variable y d (i-1) When becomes 0, the output variable y d (i)Also, a part of the output of the mapping Φ becomes 0. In this way, since the model expressed by equation (1) includes the mapping Ψ expressed by equations (8) to (11) and the mapping Φ expressed by equations (12) to (15), it is guaranteed that when a specific input becomes 0, the corresponding output becomes 0. Therefore, model learning device 200 can learn a model that can stably control the system.
[0062] <Third embodiment> 7 is a schematic diagram showing the configuration of a model learning device 300 of the third embodiment. Compared with the model learning device 100 (FIG. 1) of the first embodiment, the model learning device 300 of the third embodiment differs in the configuration of the model stored in the model storage unit.
[0063] The model learning device 300 is, for example, a personal computer (PC). The model learning device 300 is a multi-function peripheral device (PC) that includes a CPU 110, a storage unit 320, a ROM / RAM 130, a communication unit 140, and an input / output unit 150. The various units of the model learning device 300 are connected to each other via a bus. The storage unit 320 is a storage medium configured from a hard disk, a flash memory, a memory card, or the like. The storage unit 320 includes a model storage unit 321 and a dataset storage unit 122.
[0064] The model storage unit 321 pre-stores a model used for learning of a state equation by the learning unit 112. In this embodiment, the model stored in the model storage unit 321 is a state equation including a bijective mapping Ψ with an input variable v as an input, and a bijective mapping Φ with an input variable y as an input, and is defined by formula (1).
number
[0065] In this embodiment, the bijective mapping Ψ included in the formula (1) is defined by the following formulas (2) to (4) which represent a multilayer neural network.
number
number
number
[0066] In this embodiment, the bijective mapping Φ included in formula (1) is defined by formulas (12) to (15) that represent a multilayer neural network whose intermediate layer is represented by at least one of formulas (13) and (14). In this embodiment, the intermediate layer of the multilayer neural network that defines the mapping Φ alternately uses formulas (13) and (14).
number
number
number
number
[0067] The model learning method in the model learning device 300 is performed by the same method as the model learning method of the first embodiment. Specifically, first, the learning unit 112 acquires a model stored in the model storage unit 321. In this embodiment, the learning unit 112 acquires a model of formula (1) including a mapping Ψ defined by formulas (2) to (4) and a mapping Φ defined by formulas (12) to (15). After that, the learning unit 112 acquires an input / output data set, and gives an input data set to the model to estimate an output. The learning unit 112 evaluates the degree of agreement between the estimated output and the output variable data set. If the degree of agreement is not sufficient, the learning unit 112 repeats the estimation and evaluation of the output. If the degree of agreement is sufficient, the learning unit 112 ends the process.
[0068] According to the model learning device 300 of this embodiment described above, each of the mappings Ψ and Φ included in the model of equation (1) includes an exogenous input d, which is an uncontrollable input that affects the change in the output variable y. As a result, the model of equation (1) becomes a state equation that also takes into account the influence of the uncontrollable exogenous input d that affects the change in the output variable y, and by using this model, it is possible to predict the future state of the system with high accuracy. Therefore, it is possible to learn a model that can control the system with high accuracy. Moreover, the mapping Ψ represented by the multilayer neural network defined by equations (2) to (4) is obtained by dividing the input variable v (i-1) When becomes 0, the input variable v (i) too 0. That is, the output of the mapping Ψ is also 0. In addition, the mapping Φ represented by the multilayer neural network defined by equations (12) to (15) has an intermediate layer represented by at least one of equations (13) and (14). In equation (13), n c dimensional output variable y c (i-1) When becomes 0, the output variable y c (i) Also, in equation (14), n c dimensional output variable y c (i-1) , or n which is part of the output variable y d dimensional output variable y d (i-1) When becomes 0, the output variable y d (i) Also, a part of the output of the mapping Φ becomes 0. In this way, since the model expressed by equation (1) includes the mapping Ψ expressed by equations (2) to (4) and the mapping Φ expressed by equations (12) to (15), it is guaranteed that when a specific input becomes 0, the corresponding output becomes 0. Therefore, model learning device 300 can learn a model that can stably control the system.
[0069] According to the model learning device 100 of this embodiment, the mapping Ψ is expressed by a multi-layer neural network defined by equations (2) to (4). As a result, in equation (3), if all elements of the input variable v (vector) are 0, then equation (2) becomes 0. The weight W Ψ is a diagonal matrix, if the j-th element of the input variable v is 0, the j-th element of the mapping Ψ is also 0. The mapping Φ is represented by a multilayer neural network defined by equations (12) to (15). This allows, for example, c dimensional output variable y c If y is 0, c (L φ ) In other words, the component corresponding to the index c of the mapping Φ(y,d) is 0. Due to these properties, the model expressed by equation (1) has subspace invariance. Specifically, in the mapping Ψ, the weight W Ψ In the case where is a diagonal matrix, the input variable v, the internal variables u and x, and the output variable y correspond to the subscripts c and d in the formulas (12) to (15), for example, c dimensional input variable v c and d dimensional input variable v d In this case, the input variable v c When is 0, the internal variable u c Also, the time derivative x of the internal variable x is 0 when the internal variable x is 0. Therefore, the internal variable x and the internal variable u are equivalent, and x c Therefore, the output variable y a also becomes 0. In this way, in the model learned in learning unit 112, when an input component becomes 0, the corresponding output component becomes 0. Therefore, model learning device 300 can learn a model that can stably control the system.
[0070] <Fourth embodiment> 8 is a schematic diagram showing the configuration of a control device 400 according to the fourth embodiment. The control device 400 according to the fourth embodiment differs in that the CPU has a decision unit.
[0071] The control device 400 can be realized as an in-vehicle ECU (Electronic Control Unit). The control device 400 of this embodiment can be used to control the system 500. As in the first embodiment, the system 500 is, for example, an internal combustion engine, a hybrid engine, a drive engine such as a power train, a fuel cell, an induction generator, etc. The control device 400 may be, for example, a personal computer, and may be used to analyze the system 500.
[0072] The control device 400 includes a CPU 410, a storage unit 120, a ROM / RAM 130, a communication unit 140, and an input / output unit 150. The units of the control device 400 are connected to each other via a bus. At least some of the functional units of the control device 400 may be realized by an ASIC (Application Specification Integrated Circuit).
[0073] The CPU 410 includes a control unit 411, a learning unit 112, and a determination unit 413. The control unit 411, like the control unit 111 of the first embodiment, controls each unit of the control device 400 by loading a computer program stored in the ROM 130 into the RAM 130 and executing it. The learning unit 212 predicts an output variable y that represents the state of the system 500. The determining unit 413 uses the state equation learned by the learning unit 112 to determine a target value of the input variable v corresponding to the target value of the output variable y. The control unit 411 outputs the target value of the input variable v determined by the determining unit 413 to the system 500. The system 500 operates according to the target value of the input variable v that is input.
[0074] According to the control device 400 of the present embodiment described above, the model acquired by the learning unit 112 is a state equation including a bijective mapping Ψ with the input variable v input to the system 500 as an input, and a bijective mapping Φ with the output variable y output from the system 500 as an input. Such a state equation can be linearized by using the mappings Ψ and Φ as internal variables, so that it is possible to guarantee that even in a control problem using a model with a nonlinear structure, a solution of optimal control that finds an optimal input that achieves a target value of the output is unique. This allows the control device 400 to stably control the system 500.
[0075] Moreover, according to the control device 400 of this embodiment, the control unit controls the system 500 by inputting a target value of the input variable v that is determined using a model that guarantees that when a specific input becomes 0, the corresponding output becomes 0. This makes it difficult for the control device 400 to control the system 500 unintentionally. Therefore, the control device 400 can stably control the system 500.
[0076] <Modifications of this embodiment> The present invention is not limited to the above-described embodiment, and can be embodied in various forms without departing from the spirit of the present invention, for example, the following modifications are possible: In the above-described embodiment, a part of the configuration realized by hardware may be replaced by software, and conversely, a part of the configuration realized by software may be replaced by hardware.
[0077] [Variation 1] In the above-described embodiment, an example of the configuration of the model learning device or the control device including the model learning device is shown. However, the configurations of the model learning device and the control device can be modified in various ways and are not limited to these configurations. For example, at least one of the model learning device and the control device may be configured by a plurality of information processing devices (including a server device, an in-vehicle ECU, etc.) arranged on a network working together.
[0078] [Variation 2] In the above-described embodiment, an example of the procedure of the model learning method (see FIG. 2) and the predictive control method (see FIG. 4) has been shown. However, these methods can be modified in various ways and are not limited to these procedures. For example, some steps may be omitted, and other steps not described may be added. Also, the order of execution of some steps may be changed.
[0079] [Variation 3] In the above embodiment, each of the bijective mappings Ψ and Φ included in the state equation shown in formula (1) is defined by a multi-layer neural network represented by formulas (2) to (4) or formulas (5) to (7). However, the definitions of the mappings Ψ and Φ are merely examples, and they may be defined in any form. The mapping Ψ is such that at least one of the output components is 0 when at least one of the components of the input variable v is 0, and the mapping Φ is such that at least one of the output components is 0 when at least one of the components of the output variable y is 0. In this case, by using a mapping that inputs an uncontrollable exogenous input d that affects the change in the output variable y together with the internal variables, a model that can predict the future state of the system with high accuracy can be obtained.
[0080] [Variation 4] In the first embodiment, when acquiring a model stored in the model storage unit 121, the learning unit 112 defines the mapping Ψ included in the state equation of equation (1) as the internal variable u shown in equation (16) and defines the mapping Φ included in the state equation of equation (1) as the internal variable x shown in equation (17), thereby learning the state equation of equation (18) in which equation (1) is expressed by the internal variables u and x. Furthermore, in the first embodiment, the learning unit 112 reads the state equations shown in equations (19) to (21) in which equations (16) to (18) are discretized at a predetermined time step Δt at a discrete time k. These methods can also be applied to the second to fourth embodiments.
[0081] [Variation 5] In the second embodiment, the intermediate layer of the multilayer neural network that defines the mapping Ψ alternates between formula (9) and formula (10), and the intermediate layer of the multilayer neural network that defines the mapping Φ alternates between formula (13) and formula (14). In the third embodiment, the intermediate layer of the multilayer neural network that defines the mapping Φ alternates between formula (13) and formula (14). The configuration of the intermediate layer of the multilayer neural network is not limited to this. For example, the intermediate layer of the multilayer neural network that defines the mapping Ψ may be only formula (9) or only formula (10). In addition, it is not necessary to alternate between formula (9) and formula (10), but it is preferable to alternate between formula (9) and formula (10). The same applies to the intermediate layer of the multilayer neural network that defines the mapping Φ.
[0082] [Variation 6] In the fourth embodiment, a control device including the model learning device of the first embodiment has been described. The model learning device included in the control device of the fourth embodiment is not limited to this. It may be the model learning device of the second embodiment or the model learning device of the third embodiment.
[0083] Although the present aspect has been described above based on the embodiment and modified examples, the above-mentioned embodiment of the aspect is intended to facilitate understanding of the present aspect and does not limit the present aspect. The present aspect may be modified or improved without departing from the spirit and scope of the claims, and equivalents are included in the present aspect. Furthermore, if a technical feature is not described as essential in this specification, it may be deleted as appropriate.
[0084] (Application example 1) A model learning device that learns a model representing a relationship between an input variable v (vector) input to a system and an output variable y (vector) output from the system, a model storage unit that stores a model used for learning a nonlinear state equation for predicting the output variable y using the input variable v; a learning unit that learns the state equation using a model stored in the model storage unit and an input / output data set including a plurality of sets of input variable data and output variable data for the model; Equipped with The model is a state equation including a bijective mapping Ψ with the input variable v as an input and a bijective mapping Φ with the output variable y as an input, The mapping Ψ is such that when at least one component of the input variable v is 0, at least one output component is 0; The mapping Φ is such that when at least one component of the output variable y is 0, at least one output component is 0. Model learning device. (Application example 2) The model learning device according to Application Example 1, The model is defined by equation (1): The mapping Ψ is defined by Equations (2) to (4) which represent a multilayer neural network: The mapping Φ is defined by the following equations (5) to (7) which represent a multilayer neural network: Model learning device.
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
number
[0085] 100, 200, 300... Model learning device 112…Learning Department 121, 221, 321…Model memory section 400...Control device 413…Decision-making department 500…System
Claims
1. A model learning device that learns a model representing a relationship between an input variable v (vector) input to a system and an output variable y (vector) output from the system, comprising: a model storage unit configured to store a model used for learning a nonlinear state equation for predicting the output variable y using the input variable v; a learning unit that learns the state equation using a model stored in the model storage unit and an input / output data set including a plurality of sets of input variable data and output variable data for the model; Equipped with The model is a state equation including a bijective mapping Ψ having the input variable v as an input, and a bijective mapping Φ having the output variable y as an input, The mapping Ψ is such that when at least one component of the input variable v is 0, at least one output component is 0; The mapping Φ is such that when at least one component of the output variable y is 0, at least one output component is 0. Model learning device.
2. 2. The model learning device according to claim 1, The model is defined by equation (1): The mapping Ψ is defined by Equations (2) to (4) which represent a multi-layer neural network: The mapping Φ is defined by Equations (5) to (7) which represent a multi-layer neural network: Model learning device. [0010] In the above formula (1), The left side of the equal sign is the time derivative of an n-dimensional vector (n is an integer) representing the output variable y, On the right hand side of the equal sign, The input variable v is an n-dimensional vector representing controllable inputs that affect the change in the output variable y; The exogenous input d is a p-dimensional vector (p is an integer) indicating uncontrollable inputs that affect the change in the output variable y. The mapping Ψ is a function that returns an n-dimensional vector using the input variable v and the exogenous input d as input, and is a bijective mapping with respect to the input variable v, The mapping Φ is a function that returns an n-dimensional vector using the output variable y and the exogenous input d as input, and is a bijective mapping with respect to the output variable y, Function A is a function that receives the exogenous input d and returns an n × n matrix, [0025] [0030] [0045] In the above formulas (2) to (4), The subscript i represents the layer number in the multi-layer neural network. Subscript L Ψ represents the number of layers in the multilayer neural network, Weight W Ψ is expressed as an n×n matrix when the input dimension of the input variable v in the multilayer neural network is n-dimensional, Bias b Ψ is expressed by an n-dimensional vector when the input dimension of the input variable v in the multilayer neural network is n-dimensional, The activation function ψ is the first argument, W Ψ (i) (d) y (i-1) +b Ψ (i) (d) or (b) Ψ (i) (d) is a homeomorphism of [0050] [006] [0070] In the above formulas (5) to (7), The subscript i represents the layer number in the multi-layer neural network. Subscript L Φ represents the number of layers in the multilayer neural network, Weight W Φ is expressed as an n×n matrix when the input dimension of the output variable y in the multilayer neural network is n-dimensional, Bias b Φ is expressed by an n-dimensional vector when the input dimension of the output variable y in the multilayer neural network is n-dimensional, The activation function φ is the first argument, W Φ (i) (d) y (i-1) +b Φ (i) (d) or (b) Φ (i) This is a homeomorphism for (d).
3. 2. The model learning device according to claim 1, The model is defined by equation (1): The mapping Ψ is defined by equations (8) to (11) which represent a multi-layer neural network whose intermediate layer is represented by at least one of equations (9) and (10), The mapping Φ is defined by equations (12) to (15) which represent a multilayer neural network whose intermediate layer is represented by at least one of equations (13) and (14): Model learning device. [0080] In the above formula (1), The left side of the equal sign is the time derivative of an n-dimensional vector (n is an integer) representing the output variable y, On the right hand side of the equal sign, The input variable v is an n-dimensional vector representing controllable inputs that affect the change in the output variable y; The exogenous input d is a p-dimensional vector (p is an integer) indicating uncontrollable inputs that affect the change in the output variable y. The mapping Ψ is a function that returns an n-dimensional vector using the input variable v and the exogenous input d as input, and is a bijective mapping with respect to the input variable v, The mapping Φ is a function that returns an n-dimensional vector using the output variable y and the exogenous input d as input, and is a bijective mapping with respect to the output variable y, Function A is a function that receives the exogenous input d and returns an n × n matrix, [0090] [0010] ##EQU00011## ##EQU00012## In the above formulas (8) to (11), The subscript i represents the layer number in the multi-layer neural network. Subscript L Ψ represents the number of layers in the multilayer neural network, The input variable v is n a (n a is an integer) dimensional input variable v a and b (n b is an integer) dimensional input variable v b If we split it into Weight W Ψa is n a ×n a is a dimensional matrix, Bias b Ψa is n a is a dimensional vector, Weight W Ψb is n a ×n b is a dimensional matrix, Bias b Ψb is n b is a dimensional vector, The activation function ψ is the first argument, W Ψa (i) (d) v a (i-1) +b Ψa (i) , b Ψa (i) , or W Ψb (i) (d) v b (i-1) +b Ψb (i) is a homeomorphism for Function f Ψ+ is any positive-valued function (a function that returns a positive value), Function f Ψ is an arbitrary function, ##EQU00013## ##EQU00014## ##EQU00015## ##EQU00016## In the above formulas (12) to (15), The subscript i represents the layer number in the multi-layer neural network. Subscript L Φ represents the number of layers in the multilayer neural network, The output variable y is c (n c is an integer) dimensional output variable y c and d (n d is an integer) dimensional output variable y d If we split it into Weight W Φc is n c ×n c is a dimensional matrix, Bias b Φc is n c is a dimensional vector, Weight W Φd is n c ×n d is a dimensional matrix, Bias b Φd is n d is a dimensional vector, The activation function φ is the first argument, W Φc (i) (d) y c (i-1) +b Φc (i) , b Φc (i) , or W Φd (i) (d) y d (i-1) +b Φd (i) is a homeomorphism for Function f Φ+ is any positive-valued function (a function that returns a positive value), Function f Φ is an arbitrary function.
4. The model learning device according to claim 2 or 3, The learning unit defines the mapping Ψ by equation (16) using an internal variable u in the formula (1), and defines the mapping Φ by equation (17) using an internal variable x in the formula (1), thereby learning the state equation defined by equation (18). Model learning device. ##EQU00017## [0018] [0019]
5. The model learning device according to claim 4, The learning unit discretizes equation (16) at a time step of discrete time k into equation (19), and discretizes equation (17) at a time step of discrete time k into equation (20), thereby learning a state equation shown in equation (21) obtained by discretizing equation (18) at a time step of discrete time k. Model learning device. [0020] ##EQU00021## [0022] In the above formula (21), the function A′ is a matrix of the same dimension as the function A.
6. The model learning device according to claim 2 or 3, The learning unit is providing the model with the set of input variable data from the input / output data set to estimate an output; assessing the agreement between the estimated outputs and the set of output variable data from the input / output data set; learning the state equation by updating learning parameters of the model according to the evaluation results; Model learning device.
7. A control device for controlling a system, A model learning device according to any one of claims 1 to 3; a determination unit that determines a target value of the input variable v corresponding to a target value of the output variable y by using the state equation learned by the learning unit; A control unit that controls the system by inputting a target value of the input variable v determined by the determination unit. Control device.
8. A model learning method for learning a model representing a relationship between an input variable v (vector) input to a system and an output variable y (vector) output from the system by a model learning device, comprising: an acquisition step of acquiring a model used for learning a nonlinear state equation for predicting the output variable y using the input variable v; a learning step of learning the state equation using the model acquired in the acquisition step and an input / output data set including a plurality of pairs of input variable data and output variable data for the model; Equipped with The model is a state equation including a bijective mapping Ψ having the input variable v as an input, and a bijective mapping Φ having the output variable y as an input, The mapping Ψ is such that when at least one component of the input variable v is 0, at least one output component is 0; The mapping Φ is such that when at least one component of the output variable y is 0, at least one output component is 0. Model training methods.
9. A computer program that causes an information processing device to learn a model that represents a relationship between an input variable v (vector) input to a system and an output variable y (vector) output from the system, an acquisition function for acquiring a model used in learning a nonlinear state equation for predicting the output variable y using the input variable v; The state equation is learned using the model acquired by the acquisition function and an input / output data set including a plurality of sets of input variable data and output variable data for the model. and causing the information processing device to execute a learning function for learning the The model is a state equation including a bijective mapping Ψ having the input variable v as an input, and a bijective mapping Φ having the output variable y as an input, The mapping Ψ is such that when at least one component of the input variable v is 0, at least one output component is 0; The mapping Φ is such that when at least one component of the output variable y is 0, at least one output component is 0. Computer program.
Citation Information
Patent Citations
Reinforcement learning methods, reinforcement learning programs, and reinforcement learning systems
JP2020144484A
Model learning device, model learning method, control device, control method, and computer program
JP2021189934A
Device, method, and program
JP2022016709A
Model learning device, controller, model learning method and computer program
JP2022064643A
Model learning device, control device, model learning method, and computer program
JP2022143499A
