JPH07210207A - Adaptive controller - Google Patents

Abstract

(57) [Summary] [Purpose] A linear control system can be constructed based on a linear model without considering the effects of non-linearity and disturbance of the controlled object, and a stable response can be obtained. [Structure] The neural network 4 outputs the output signal Ubf (k) 16 of the compensator at the present time and the output signals [Y (k), Y (k-1), ...
Y (k−n)] and manipulated variables [U (k−1), U (k−2), ...,
[U (k−m)] is input and the operation amount U (k) 3 is output. The learning means 8 learns the neural network 4 so that the error between the output 2 of the controlled object and the output 7 of the linear model is minimized.

Description

Detailed Description of the Invention

[0001]

The present invention relates to a robot, a machine tool,
The present invention relates to an adaptive control device using a neural network that realizes a highly accurate positioning control system that can cope with nonlinearity of a controlled object, disturbance, and environmental change over time in a positioning control device such as an XY stage and a disk device.

[0002]

2. Description of the Related Art Adaptive control systems include model reference adaptive control and self-tuning regulators. These methods are effective only when the dynamic characteristic model of the controlled object can be described by linear dynamics, but cannot deal with the nonlinearity of the controlled object, disturbance, and environmental change over time.

On the other hand, it is known that a neural network has a relatively simple structure and is capable of learning a dynamics model including a non-linear system. As a means to deal with the non-linearity of the controlled object and the environmental change over time,
Japanese Unexamined Patent Publication No. 2-54304 discloses an adaptive control system using a neural network.

FIG. 8 shows the adaptive control device disclosed therein. This is a device in which an addition point 61 is provided in a feedback loop including a controlled object 51 and a feedback compensator 57, and a feedforward loop from a target value 59 to the addition point is configured by a neural network 54.

The operation of this device is as follows. That is, before the neural network starts learning,
The target value R59 is the output Y of the control target for each control sampling.
52, the feedback compensator 57 calculates based on the deviation 60, and the calculation result Ub58 is applied to the controlled object 51 as the manipulated variable U53. When the learning of the neural network 54 is started, the target value, the velocity target value that is the first derivative thereof and the acceleration target value that is the second derivative thereof are input to the neural network, and the output Uf55 of the neural network is output as the output Ub5 of the feedback compensator.
Then, the manipulated variable U53 is applied to the controlled object 51.

In the figure, s is a differential operator. The neural network 54 uses the manipulated variable U53 as a teacher signal, and the learning proceeds so as to minimize J = (U−Uf) ² = Ub ² . That is, as the learning progresses, the output Ub58 of the feedback compensator 57 becomes smaller, the inverse model of the controlled object 51 is gradually constructed in the neural network 54, and the feedforward control by the obtained inverse model is started. The back propagation method is used for learning of the neural network. Since learning can be performed while controlling, and the manipulated variable can be adaptively calculated with respect to the characteristic variation and non-linearity of the controlled object, its application to a robot or the like is being studied.

[0007]

However, in the prior art, since the neural network is used as the feedforward compensator, the followability with respect to the target value is improved, but the effect of suppressing the disturbance or noise added to the controlled object is improved. There was a problem that it was not enough. In particular, in the conventional technology, since only the target value, the velocity target value, and the acceleration target value are input as the input signal of the neural network,
It was difficult to construct the inverse dynamics of the controlled object, which is affected by dynamics fluctuation and disturbance, in the neural network.

Further, in the conventional technique, there is a problem that it is impossible to effectively utilize the information of the linear dynamics of the non-linear system which can be obtained in advance. In particular, the conventional technique requires a feedback compensator that stabilizes a nonlinear control target. Normally, when designing a feedback compensator, linearized dynamics information of a non-linear control target is required, but in the prior art, this linearized dynamics information is effectively used in learning a neural network. Was difficult.

An object of the present invention is to use a neural network which has a sufficient effect of suppressing non-linearity and disturbance of a controlled object, reduces the learning load of the neural network, and is capable of highly accurate learning. An object is to provide an adaptive control device.

[0010]

A first invention for achieving the above object is to feedback-control a control deviation between a target value and the control output so that a control output of a controlled object coincides with the target value. In a control device having a controller, a neural network model having an output of the controller as an input signal, a linear model having an output of the controller as an input signal, and an output of the neural network as an input signal of the controlled object. , Learning means for changing the weight between the layers of the neural network so that the output of the controlled object and the output of the linear model become equal.

A second aspect of the present invention for achieving the above object has a controller for feedback-controlling the control deviation between the target value and the control output so that the control output of the controlled object coincides with the target value. In the control device, a neural network model having an output of the controller as an input signal, a linear model having an output of the controller as an input signal, and an output of the controller as an input signal of the control target,
Learning means for changing the weights between the layers of the neural network so that the output of the controlled object and the signal of the output of the neural network added to the output of the linear model are equal, and the output of the neural network is And means for correcting the output of the controlled object.

Further, a third invention for achieving the above object has a controller for performing feedback control of a control deviation between the target value and the control output so that the control output of the controlled object coincides with the target value. In the control device, a neural network model for inputting an input signal to the controlled object, a signal obtained by adding an output of the neural network and an input signal to the controlled object, and a linear model having the added signal as an input signal And learning means for changing the weight between the layers of the neural network so that the output of the controlled object and the output of the linear model become equal, and a means for correcting the target value by the output of the neural network. .

Further, when the first invention is applied to a disc device, a disc on which information is recorded, a head for reading or writing information between the disc and the disc is recorded. A head position detector that detects a deviation between the position information and the position of the head, and outputs a deviation position signal;
A controller comprising: a controller for feedback-controlling the deviation position signal so as to make the deviation position signal zero; and an actuator that moves integrally with the head using the output of the controller as a drive input, A neural network model whose input signal is the output of, a linear model whose output is the output of the controller, an output of the neural network is the input signal of the actuator, and the deviation position signal and the output of the linear model are And learning means for changing the weights between the layers of the neural network so as to be equal.

When the second invention is applied to a disk device, a disc on which information is recorded, a head for reading or writing information between the disc, and the disc. A head position detector that detects a deviation between the recorded position information and the position of the head and outputs a deviation position signal, and a controller that feedback-controls the deviation position signal so that the deviation position signal becomes zero. In a disk device including an actuator that moves integrally with the head that uses the output of the controller as a drive input, a neural network model that uses the output of the controller as an input signal and an output of the controller as an input signal A linear model,
The output of the controller is used as the input signal of the actuator, and the deviation position signal and the weight of each layer of the neural network are changed so that the signal obtained by adding the output of the neural network to the output of the linear model becomes equal. The learning means and the output of the neural network correct the deviation position signal.

When the third invention is applied to a disk device, a disc on which information is recorded, a head for reading or writing information between the disc, and the disc. A head position detector that detects a deviation between the recorded position information and the position of the head and outputs a deviation position signal, and a controller that feedback-controls the deviation position signal so that the deviation position signal becomes zero. In a disk device including an actuator that moves integrally with the head, which uses the output of the controller as a drive input, a neural network model that inputs an input signal to the actuator, an output of the neural network and an input to the actuator A signal and a linear model having the added signal as an input signal; And learning means for output of the linear model to change the weight of each layer of the neural network to be equal, the output of the neural network is intended to include a means for correcting the output of the controller.

[0016]

In the present invention, since the manipulated variable and the target value are corrected by using the neural network so that the controlled object becomes a desired linear model, it is set without considering the influence of the non-linearity or the disturbance of the controlled object. Various linear control systems can be constructed based on the linear model, and high-accuracy and high-speed response can be obtained. The desired linear model set by the designer is preferably set to the linearized dynamics around the equilibrium point of the nonlinear controlled object, but the error between the set linear model and the actual linearized dynamics is It is not necessary to set a particularly exact linearization model as it is captured and properly compensated in the neural network.

Further, in the first invention, since the neural network is inserted in series in the control loop, it is particularly effective when multiplicative nonlinearity or disturbance acts on the controlled object. In the second invention, since the neural network corrects the target value additively, it is particularly effective when additive nonlinearity or disturbance acts on the output end of the controlled object. In the third invention, since the neural network corrects the manipulated variable additively, it is particularly effective when additive nonlinearity or disturbance acts on the input end of the controlled object.

As described above, the present invention provides means for effectively correcting by using a neural network in accordance with a non-linearity of a controlled object or a location where disturbance or the like acts. Not only the influence of the non-linearity and the disturbance is suppressed, but also the influence of the non-linearity and the disturbance acting in other places can be effectively suppressed.

[0019]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of the present invention will be described in detail below with reference to the drawings.

FIG. 1 is a diagram showing an embodiment of the first invention. The first invention is particularly effective when multiplicative non-linearity or disturbance acts on the controlled object.

In FIG. 1, 1 is a controlled object, 2 is an output Y of the controlled object, 3 is an input U to the controlled object, 4 is a neural network, 6 is a linear model specified by the designer, and 8 is an error signal 9. Learning means for computing the connection weight between the layers of the neural network 4 so as to minimize it by using a backpropagation method, 14 is a target value R, and 10 is feedback designed so that the linear model 6 can be stably controlled. The compensator 12 is a feedforward compensator designed for high response based on the inverse dynamics of the linear model 6. Here, the linear model 6 is preferably set to a model in which a non-linear control target is linearized around the equilibrium point, but the linear model 6 is not limited to this and may be a desired linear model. When it is difficult to derive the physical model, for example, a second-order linear model is assumed, parameter identification around the equilibrium point is performed in advance, and the obtained result may be used as the linear model. Further, the feedforward compensator 12 need not be used unless high responsiveness is required.

Next, the flow of control at time k in this embodiment will be described. First, the output Y of the controlled object
(k) Read 2. The feedback compensator 10 generates Ub (k) 11 based on the deviation between the target value R (k) 14 and the output Y (k) 2 of the controlled object. The feedforward compensator 12 generates Uf (k) 13 based on the target value R (k) 14. Uf (k) 13 and Ub (k) 11 are added, and Ubf
(k) It becomes 16. Neural network 4 is Ubf
(k) 16 and the output Y (k) of the controlled object, the output Y (k-1) of one step past, ..., The output Y (k of the past n steps
-N), and the operation amount U (k-1), 2 one step past
The operation amount U (k−2), which is past step, is set to the input layer. The selection of steps n and m is preferably set to be equal to or higher than the orders of the denominator and the numerator of the linear model 6.

The neural network 4 preferably has three or more layers for the purpose of acquiring non-linearity and disturbance added to the controlled object, but it is also possible to reduce the arithmetic processing by using two layers of linear neurons. Alternatively, a recurrent neural network that feeds back the output of the intermediate layer to the input of the input layer may be used. In this case, the output of the controlled object Y (k) 2, the manipulated variable U (k-1) 3, the output of the compensator U
Only bf (k) 16 needs to be input to the input layer of the neural network 4. The output of the neural network is k
The operation amount U (k) 3 for the controlled object in the step is obtained.

After the above calculation is completed, learning of the neural network 4 is performed. In the present invention, the product of the dynamics of the neural network 4 and the controlled object 1 is
The neural network 4 is trained so as to be equal to the set linear model 6. That is, using the learning means 8, the output Y (k) 2 of the controlled object and the output Ym of the linear model are obtained.
(k) Learning of the neural network 4 is performed so that the difference of 7 is minimized.

First, the structure of the neural network 4 will be described. The structure of the neural network basically has a multilayer structure. For example, in the three-layer structure, each neuron in the input layer becomes an input of a neuron in the intermediate layer,
The output of the neuron in the middle layer becomes the input of the output layer. Hereinafter, the number of layers will be limited to three for the description. Here, the coupling load between the input layer and the intermediate layer is Wij, and the coupling load between the intermediate layer and the output layer is Vj. The input / output characteristics of each neuron are modeled by a sigmoid function. The sigmoid function F (X) is

[0026]

## EQU1 ## F (X) = 2 / (1 + exp (-X))-1 (Expression 1) Here, the neurons in the input layer do not undergo sigmoid transformation, and the neurons in the intermediate layer and the output layer do sigmoid transformation. The input signal to the input layer is Xi (k) (i
= 1 to n + m + 2), the output signal of the intermediate layer is Hj (k) (j
= 1 to q), and the output signal of the output layer is U (k) (equal to the manipulated variable U (k) in the first invention). However, the number of neurons q in the intermediate layer needs to be determined by trial and error. Then, Hj (k) and U (k) are as shown in Equations 2 and 3, respectively. However, the offset amount is not described here.

[0027]

[Equation 2]

[0028]

[Equation 3]

Next, the learning algorithm of the neural network 4 will be described. The learning means 8 of the neural network 4 is performed by changing the connection weight between neurons. Here, the back propagation method (back propagation algorithm) will be described. The learning unit 8 squares the error E (k) between the output Y (k) 2 of the controlled object and the output Ym (k) 7 of the linear model, that is,

[0030]

Equation 4] J (k) = (Y ( k) -Ym (k)) a ^{2/2 = E (k)} 2/2 ... ( Equation 4) so as to minimize, performs varying the connection weights . This learning process is sequentially performed from the output layer to the input layer.

The coupling load is changed according to the equations (5) to (8).

[0032]

[Formula 5] ΔVj (k) = − η ・ ∂J (k) / ∂Vj (k) = η ・ E (k) ・ ∂Y (k) / ∂U (k) ・ (1-U (k ) ² ) / 2 · Hj (k) (Equation 5)

[0033]

[Equation 6] ΔWji (k) = − η ・ ∂J (k) / ∂Wji (k) = η ・ E (k) ・ ∂Y (k) / ∂U (k) ・ (1-U (k ) ² ) ・ Vj (k) ・ (1-Hj (k) ² ) / 2 ・ Xi (k) (Equation 6)

[0034]

[Expression 7] Vj (k + 1) = Vj (k) + ΔVj (k) (Expression 7)

[0035]

[Equation 8] Wji (k + 1) = Wji (k) + ΔWji (k) (Equation 8) where η is a learning constant, which is empirically determined from the number of neurons, the number of layers, and input / output values. To be done. In addition, in the first invention, ∂Y (k) / ∂U is used in the calculation of Formula 5 and Formula 6.
Information on the controlled object in (k) is required. If the non-linearity of the controlled object is weak, the DC gain of the linearized model may be set to ∂Y (k) / ∂U (k). When the nonlinearity of the controlled object is strong, the dynamics of the controlled object are learned in advance by another neural network for identification, and input from the identified neural network instead of ∂Y (k) / ∂U (k). It suffices to calculate the change in the output with respect to the change in.

FIG. 2 shows an adaptive control device using a neural network in which the neural networks connected in series in the control loop in FIG. 1 are arranged in parallel with the control loop. Here, 17 is an addition point provided in the loop. The output Un of the neural network 4 additively corrects the output signal Ubf16 of the compensator. That is, in the present embodiment, the characteristics can be improved by adding the neural network 4, the learning means 8 and the linear model 6 without changing the structure of the conventional control system.
Furthermore, the correction of the output of the neural network into the loop can be controlled based on the convergence of the error signal E (k) 9.

FIG. 3 is a block diagram showing an embodiment of the second invention. The second invention is particularly effective when additive nonlinearity or disturbance acts on the output end of the controlled object.

The control flow at the time of step k will be described. First, the output Y (k) 2 to be controlled is read. The neural network 4 outputs the output Y of the controlled object.
(k), one step past output Y (k−1), ..., n steps past output Y (k−n), and one step past manipulated variable U (k−1), two steps past manipulated variable U (k-2),
..., m operation steps past U (km) are set in the input layer. Next, output Yn (k) 5 of the neural network
Is used to correct the target value R (k) 14. The feedback compensator 10 generates Ub (k) 11 based on the deviation between the corrected target value 18 and the output 2 to be controlled, and the feedforward compensator 12 generates Uf (k) 11 based on the corrected target value 18. (k) 13 is generated. Uf (k) 13 and Ub (k) 1
1 is added and becomes the manipulated variable U (k) 3 at the k step.

After the above calculation is completed, learning of the neural network 4 is performed. In the present invention, the output 5 of the neural network and the output 7 of the set linear model are set.
Learning of the neural network is performed so that the added signal of is equal to the output 2 of the controlled object. That is, the learning means 8 is a squared error of the error E (k) between the output Y (k) 2 of the controlled object, the output Ym (k) 7 of the linear model, and the output Yn (k) of the neural network, that is,

[0040]

Equation 9] J (k) = a (Y (k) -Ym (k ) -Yn (k)) 2/2 = E (k) 2/2 ... ( number 9) so as to minimize coupling load To change. This learning process is sequentially performed from the output layer to the input layer.

The coupling load is changed according to the following equation.

[0042]

[Formula 10] ΔVj (k) = − η · ∂J (k) / ∂Vj (k) = η · E (k) · (1-Yn (k) ² ) / 2 · Hj (k)… ( (Number 10)

[0043]

[Expression 11] ΔWji (k) = − η · ∂J (k) / ∂Wji (k) = η · E (k) · (1-Yn (k) ² ) · Vj (k) · (1- Hj (k) ² ) / 2 · Xi (k) (Equation 11)

[0044]

[Equation 12] Vj (k + 1) = Vj (k) + ΔVj (k) (Equation 12)

[0045]

[Equation 13] Wji (k + 1) = Wji (k) + ΔWji (k) (Equation 13) From the above equation, in the second invention, it is necessary to calculate Equations 5 and 6 in the first invention. It does not require the prior information of the controlled object of ∂Y (k) / ∂U (k).

FIG. 4 is a block diagram showing an embodiment of the third invention. The third invention is particularly effective when additive non-linearity or disturbance acts on the input end of the controlled object.

The control flow at the time of step k will be described. First, the output Y (k) 2 to be controlled is read. The feedback compensator 10 generates Ub (k) 11 based on the deviation between the target value R (k) 14 and the output 2 of the controlled object. The feedforward compensator 12 has a target value R
Uf (k) 13 is generated based on (k) 14. The neural network 4 outputs the output Y (k) of the controlled object, the output Y (k-1) of one step past, ..., The output Y of n steps past.
(k−n), and the operation amount U (k−1) one step past,
The operation amount U (k−2), which is two steps past, is set to the input layer, the operation amount U (k−m) which is m steps past. The output Yn (k) 5 of the neural network 4 is Uf (k) 13 and Ub
The signal added by (k) 11 is corrected to obtain the manipulated variable U (k) 3 at the k step.

After the above calculation is completed, learning of the neural network 4 is performed. In the present invention, the output Ym (k) 7 of the linear model 4 to which the output signal Yn (k) 5 of the neural network and the addition signal of the manipulated variable U (k) 3 are input is equal to the output Y (k) 2 of the controlled object. Learn the neural network so that That is, the learning means 8 has a squared error of the error E (k) between the output Y (k) 2 of the controlled object and the output Ym (k) 7 of the linear model, that is,

[0049]

Equation 14] J (k) = (Y ( k) -Ym (k)) a ^{2/2 = E (k)} 2/2 ... ( number 14) to minimize, performs varying the connection weights . This learning process is sequentially performed from the output layer to the input layer.

The coupling load is changed according to the following equation.

[0051]

[Formula 15] ΔVj (k) = − η · ∂J (k) / ∂Vj (k) = η · E (k) · b0 · (1-Yn (k) ² ) / 2 · Hj (k) … (Equation 15)

[0052]

[Expression 16] ΔWji (k) = − η · ∂J (k) / ∂Wji (k) = η · E (k) · b0 · (1-Yn (k) ² ) · Vj (k) · ( 1-Hj (k) ² ) / 2 · Xi (k) (Equation 16)

[0053]

[Expression 17] Vj (k + 1) = Vj (k) + ΔVj (k) (Expression 17)

[0054]

Wji (k + 1) = Wji (k) + ΔWji (k) (Equation 18) where b0 is the DC gain of the linear model 6, that is,
b0 = ∂Ym (k) / ∂Yn (k). From the above equation, the third invention does not require the prior information of the control target of ∂Y (k) / ∂U (k), which was required when calculating the equations 5 and 6 in the first invention.

FIG. 5 is an example of a block diagram of hardware for realizing the above-described embodiment of the first invention in a disk device. The configuration method shown in FIG. 5 is also effective as a hardware configuration method for realizing the above-described embodiments of the second invention and the third invention.

As shown in FIG. 5, the controlled object 1 in FIG. 1 comprises a head support mechanism system. The head support mechanism system
Specifically, a mechanism 32 that supports the head 31, a voice coil motor 33 that drives the mechanism, a power amplifier 34 that applies a current to the voice coil motor, a DA converter 35 that converts a digital signal into an analog signal, and a disk circle. A head position detecting means 38 for detecting the deviation position signal 37: Ye between the target track 36 on the track recorded on the plate 30 and the head 31, and an AD converter 39 for converting an analog signal into a digital signal are DA converters. To the AD converter.

These dynamics are basically modeled by a quadratic delay system. However, the gain variation due to the position of the voice coil motor 33 and the head position detecting means 3
8 and a multiplicative non-linear variation such as dynamic variation due to the temperature of the head support mechanism 32. Further, the target trajectory 36 recorded on the track is slightly undulated from a perfect circle due to the vibration of the disk device at the time of recording the target trajectory and the rotational vibration of the disk disc 30. That is, in the disk device, there is an additive disturbance at the output end of the controlled object. The control purpose of the disk device is to eliminate the influence of multiplicative / additive non-linearity applied to the controlled object, and to make the deviation between the deviation position signal and the target value zero. It is solved by the second and third inventions.

The arithmetic processing in FIG. 5 is executed in the microprocessor system 46. In the microprocessor system 46, the microprocessor 41 and the neuroprocessor 40 are connected to a RAM (random access memory) 42 and a ROM (read-only memory) 43 via a bus line 45. The ROM 43 stores the structure of the neural network 4, the learning means 8, the feedback compensator 10, the feedforward compensator 12, and the program of the linear model 6. The RAM 42 plays a role of temporarily storing an input / output signal to be controlled.
The neuroprocessor 40 may be replaced by the microprocessor 41, but the neural network operation (Equations 2 and 3) and learning operation (Equations 7 and 8) are performed at high speed based on the output result of the AD converter 39. To run. The output of the neural network is the manipulated variable 3, and the bus line 4
It is applied to the DA converter 35 via 5. The command generator 44 monitors the output signal of the AD converter, and the learning means 9
Play the role of a switch.

FIG. 6 is an example of a gain diagram of the head support mechanism system from the operation amount U3 of the disk device to the output Y2 of the controlled object when the above embodiment is not used. The results of measurement three times while changing the gain of the excitation signal are plotted.
Basically, it can be modeled as a second-order lag system, but it can be seen that the influence of viscous resistance of bearings etc. in the low range and the influence of the support system supporting the head appear in the high range. Conventionally, each compensator has been designed by modeling the controlled object with an averaged second-order lag system without considering these effects.

On the other hand, FIG. 7 is an example of a gain diagram of the head support mechanism system from the output signal Ubf16 of the compensator to the output Y2 to be controlled when the embodiment is used. Similar to FIG. 6, the results of measurement three times while changing the gain of the excitation signal are plotted. It can be seen that the neural network compensates the low-frequency nonlinearity and the high-frequency nonlinearity, and the linear model of the second-order lag system is set by the designer in all frequency bands. As a result, the feedback compensator and the feedforward compensator based on the linear model work effectively, and it becomes possible to obtain a stable response even if the controlled object has nonlinearity or disturbance.

Although the disk device has been described in the embodiment, the same effects as described above can be obtained by applying the present invention to a system of other mechatronics equipment such as a robot.

[0062]

According to the present invention, the manipulated variable and the target value are corrected by using a neural network so that the controlled object becomes a desired linear model. Therefore, the linear control system can be configured based on the set linear model without considering the influence of the non-linearity of the controlled object or the disturbance, and the stable response can be obtained.

[Brief description of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention.

FIG. 2 is a block diagram showing a second embodiment of the present invention.

FIG. 3 is a block diagram showing a third embodiment of the present invention.

FIG. 4 is a block diagram showing a fourth embodiment of the present invention.

FIG. 5 is a block diagram showing an example of a disk device of the present invention.

FIG. 6 is a gain diagram when the present invention is not used.

FIG. 7 is a gain diagram when the present invention is used.

FIG. 8 is a conventional block diagram.

[Explanation of symbols]

1 ... Control object, 2 ... Output of control object, 3 ... Operation amount, 4 ...
Neural network, 6 ... Linear model, 7 ... Linear model output, 8 ... Learning means, 9 ... Error signal to learning means, 10 ... Feedback compensator, 11 ... Feedback compensator output, 12 ... Feedforward compensator, 1
3 ... Output of feedforward compensator, 14 ... Target value,
15 ... Deviation between target value and output of controlled object.

Claims (1)

[Claims] 1. A controller having a controller for feedback-controlling a control deviation between the target value and the control output so that a control output of a controlled object coincides with the target value, wherein an output of the controller is an input signal. A neural network model, a linear model using the output of the controller as an input signal, and an output of the neural network as the input signal of the controlled object so that the output of the controlled object and the output of the linear model become equal. An adaptive control device comprising: learning means for changing a weight between layers of the neural network.