JP2018147103A - Model learning device, controlled variable calculation device, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To obtain control rules for controlling an automatic control vehicle in a traffic environment including a manual control vehicle and an automatic control vehicle.SOLUTION: On the basis of multiple pieces of learning data that are collected beforehand and represent state parameters of manual control vehicles and state parameters of automatic control vehicles, for each state space representing states of multiple vehicles including the manual control vehicles and the automatic control vehicles, a state space model learning part 122 learns a model for predicting a variation in the state variables of the multiple vehicles. On the basis of the model that is learnt by the state space model learning part 122 regarding the state space, for each state space, a control rule design part 124 then generates control rules for controlling the state parameter of each of the automatic control vehicles in the multiple vehicles.SELECTED DRAWING: Figure 1

Description

  The present invention relates to a model learning device, a control amount calculation device, and a program.

  2. Description of the Related Art Conventionally, a lane change optimization device that performs control of a host vehicle using model predictive control while predicting the behavior of surrounding vehicles is known (Patent Document 1).

  Further, a method is known in which a state space equation based on the formation of a plurality of moving bodies is defined, and a control law for performing collision avoidance using the equation is designed off-line (Non-Patent Document 1).

  In addition, a method is known in which a control input that allows a plurality of vehicles to smoothly merge in an merging scene is obtained and executed as an optimal control law (Non-Patent Document 2).

JP 2006-209455 A

Miyazaki Tatsuya, Takaba Joruri, “Formation Control of Mobile Robots Considering Obstacle Avoidance”, Journal of System Control Information Society, 2015 J. Rios-Torres, A.A. Mailkopoulos and P. Pisu, "Online Optimal Control of Connected Vehicles for Efficient Traffic Flow at Merging Roads," in Proc. Of the IEEE 18th International Conference on Intelligent Transportation Systems, 2015.

  However, since Patent Document 1 is a method using model predictive control, there arises a problem that the real-time property is lost when the number of vehicles increases.

  In Non-Patent Document 1, real-time performance is high due to off-line control design, but it is premised on that all moving bodies can be controlled, and it is difficult to control if a non-predictable moving body such as a manually operated vehicle is included. The problem arises.

  Non-Patent Document 2 is based on an optimal control framework, so control performance with respect to a desired target is expected. However, it is assumed that all vehicles can be controlled. When a difficult moving body is included, there arises a problem that it is difficult to control.

  The present invention has been made to solve the above-mentioned problems, and it is possible to obtain a control law for controlling an automatically controlled moving body in a traffic environment including a manually controlled moving body and an automatically controlled moving body. An object of the present invention is to provide a model learning apparatus and program that can be used.

  It is another object of the present invention to provide a control amount calculation apparatus and program capable of obtaining a control amount for controlling an automatic control mobile body in a traffic environment including a manual control mobile body and an automatic control mobile body.

  In order to achieve the above object, the model learning device of the present invention is configured to detect the states of a plurality of moving bodies including a manually controlled moving body that is a moving body by manual manual control and an automatically controlled moving body that is a moving body by automatic control. For each state space to represent, the amount of change in the state variables of the plurality of moving bodies is predicted based on a plurality of learning data representing the state variables of the manually controlled moving body and the state variables of the automatic control moving body that are collected in advance. A learning unit for learning a model for performing each of the automatic control moving bodies of the plurality of moving bodies based on the model learned for the state space by the learning unit for each state space And a control law design unit that generates a control law for controlling the state variables.

  Further, the program of the present invention is a computer for each state space that represents the state of a plurality of moving bodies including a manually controlled moving body that is a manually controlled moving body and an automatically controlled moving body that is a automatically controlled moving body. Learning a model for predicting the amount of change in the state variables of the plurality of moving bodies based on a plurality of learning data representing the state variables of the manually controlled moving bodies and the state variables of the automatic control moving bodies collected in advance. And for each state space, to control each state variable of the automatically controlled mobile body among the plurality of mobile bodies based on the model learned about the state space by the learning section This is a program for functioning as a control law design unit that generates the control law.

  According to the present invention, the learning unit performs in advance for each state space representing the states of a plurality of moving bodies including a manually controlled moving body that is a manually controlled moving body and an automatically controlled moving body that is a automatically controlled moving body. A model for predicting the amount of change in the state variables of the plurality of mobile bodies is learned based on the collected plurality of learning data representing the state variables of the manually controlled mobile bodies and the state variables of the automatically controlled mobile bodies. Also, a control law for controlling each state variable of the automatically controlled mobile body among the plurality of mobile bodies based on the model learned about the state space by the learning section for each state space by the control law design section. Is generated.

  In this way, for each state space representing the states of the plurality of moving bodies, a plurality of moving bodies are obtained based on the plurality of learning data representing the state variables of the manually controlled moving bodies and the state variables of the automatic control moving bodies that are collected in advance. Learning a model for predicting the amount of change in state variables, and generating a control law for an automatically controlled mobile body among multiple mobile bodies based on the model learned for the state space for each state space Thus, it is possible to obtain a control law for controlling the automatic control moving body in a traffic environment including the manual control moving body and the automatic control moving body.

  Further, the learning unit of the present invention, as the learning data, the state variable of the moving body, the state variable of the moving body around the moving body, and the state of the moving body for each of the moving bodies The model can be learned using the amount of change of the variable.

  In addition, the control law design unit may include the model learned by the learning unit and the control law for optimizing an evaluation function expressed using the control law, a power polynomial of the state variable, the state variable The control law can be generated by obtaining the Chebyshev polynomial or the HJB equation using the neural network having the state variable as an input as a basis function.

  Further, the learning unit receives a plurality of state variables of the moving bodies as inputs, and outputs an average of the amount of change of the state variables of the plurality of moving bodies and a variance of the amount of change of the state variables of the plurality of moving bodies. A Gaussian process model to be learned can be learned as the model.

  Further, the state variable may include a position and a speed of the moving body.

  The control amount calculation apparatus according to the present invention acquires a state variable for each of a plurality of moving bodies and information on which of the manually controlled moving body and the automatic control moving body, and the plurality of moving bodies belong to A state acquisition unit that specifies a state space; a control law acquisition unit that acquires the control law generated by the model learning device according to the state space specified by the state acquisition unit; and the plurality of movements Based on the state variable of the body and the control law acquired by the control law acquisition unit, a control amount calculation unit that calculates the control amount of the state variable for the automatic control mobile body among the plurality of mobile bodies, It is comprised including. Thereby, even in a traffic environment including a manually controlled moving body and an automatically controlled moving body, a control amount for controlling the automatically controlled moving body can be obtained.

  As described above, according to the model learning device and the program of the present invention, the state variable of the manually controlled moving body and the state variable of the automatically controlled moving body collected in advance for each state space representing the state of the plurality of moving bodies. A model for predicting the amount of change in the state variables of a plurality of mobile objects is learned based on a plurality of learning data representing a plurality of mobile objects based on a model learned for the state space for each state space. It is possible to obtain a control law for controlling an automatic control mobile body in a traffic environment including a manual control mobile body and an automatic control mobile body by generating a control law for the automatic control mobile body An effect is obtained.

  According to the control amount calculation device and the program of the present invention, the state variable for each of the plurality of moving bodies and the information on which of the manual control moving body and the automatic control moving body are acquired and the plurality of movements The state space to which the body belongs is specified, the control law generated by the control law generation device is acquired according to the specified state space, and based on the state variables of the plurality of moving bodies and the acquired control law By calculating a control amount for an automatically controlled mobile body among a plurality of mobile bodies, a control amount for controlling the automatically controlled mobile body can be obtained even in a traffic environment including a manually controlled mobile body and an automatically controlled mobile body. The effect that it can obtain is acquired.

It is a block diagram which shows the structure of the vehicle control system of embodiment of this invention. It is a flowchart which shows the content of the learning process routine in the control law production | generation apparatus of embodiment of this invention. It is a flowchart which shows the content of the control law design processing routine in the control law production | generation apparatus of embodiment of this invention. It is a flowchart which shows the content of the control amount calculation process routine in the control amount calculation apparatus of embodiment of this invention.

  Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings. In the embodiment of the present invention, a case where a vehicle which is an example of a moving body is targeted will be described as an example.

<Configuration of vehicle control system>
A vehicle control system according to the present embodiment will be described. As shown in FIG. 1, a vehicle control system 10 according to the present embodiment includes a control law generation device 12 that generates a control law for controlling a vehicle based on learning data collected in advance, and a control law generation And a vehicle control amount calculation device 14 that calculates a control amount for the vehicle based on the control law generated by the device 12. The control law generation device 12 and the vehicle control amount calculation device 14 are connected by communication means such as the Internet. The control law generation device 12 is an example of a model learning device.

[Control law generator]

  The control law generation device 12 includes a CPU that controls the control law generation device 12 as a whole, a ROM as a storage medium that stores a program of each processing routine, a RAM that temporarily stores data as a work area, and a bus that connects these It is comprised by the computer containing. In the case of such a configuration, a program for realizing the function of each component is stored in a storage medium such as a ROM or HDD, and each function is realized by executing the program by the CPU. To.

  If this computer is described in terms of functional blocks divided for each function realizing means determined based on hardware and software, the control law generating device 12 includes a learning data collecting unit 118, a learning database, as shown in FIG. 120, a state space model learning unit 122, a control law design unit 124, a control law database 126, a control unit 128, and a communication unit 130. The state space model learning unit 122 is an example of a learning unit.

  In the present embodiment, the amount of change in the state variables of the plurality of vehicles is predicted for each state space representing the state of the plurality of vehicles including the manually controlled vehicle that is a manually controlled vehicle and the automatically controlled vehicle that is a automatically controlled vehicle. Learn the model to do. The state space of the present embodiment will be described below.

In the following, the i-th component of the vector vεR ⁿ is denoted as [v] _i . In addition, an i row and j column component of the matrix MεR ^{n × m} is denoted as [M] _{i, j} . Further, the matrix M∈R ^{n × m} i row vector and j column vector each _{[M] i, ·, [} M] ·, denoted by _j.

In this embodiment, a case where control is performed in the front-rear direction of the automatic control vehicle will be described as an example. First, a state space is defined for N vehicles. The absolute position in the front-rear direction of the vehicle i at time t is p _{glo, i} (t), and the absolute speed is v _{glo, i} (t). A steady state related to the target positions of the N vehicles is set to p _{ref, i} (t), and a steady state related to the target speed is set to v _{ref, i} (t). i represents an index for identifying the vehicle.

Further, the relative position from the target steady state is set as p _i : = p _{glo, i} −p _{ref, i} (t). In addition, the target relative speed from the steady state is set as v _i : = v _{glo, i} −v _{ref, i} (t). In this case, the state variable x (t) is defined by the following equation (1).

(1)

  As shown in the above formula (1), in the present embodiment, the position of the vehicle and the speed of the vehicle are included as state variables. The state space is a 2N-dimensional Euclidean space.

Next, for M vehicles, when M vehicles i ₁ , i ₂ ,..., I _M can be controlled as automatic control vehicles, the state space equation is given by the following equation (2): It is done.

(2)

Here, the formula (2) u (t) in ∈R ^M is a control amount for automatic control vehicle M stage (e.g., the acceleration given to the automatic control vehicle). In the present embodiment, the passive dynamics f (x (t)) R ^2N in the above equation (2) is modeled by a Gaussian process. Further, in the matrix B in the above formula (2), each element shown in the following formula (A) is a non-zero constant (coefficient for acceleration), and the values of the other elements are zero.

(A)

In the above formula (A), for example, when the second vehicle is a manually controlled vehicle and the first vehicle and the third vehicle are automatically controlled vehicles, i ₁ = 1 and i ₂ = 3. . In each element of the matrix B shown in the above equation (2), coefficients for the control amounts of speed and acceleration for each vehicle are arranged and stored. Therefore, a coefficient for each vehicle is arranged in order, such as a coefficient for the speed of the first vehicle 1, a coefficient for the acceleration of the first vehicle, a coefficient for the speed of the second vehicle, and a coefficient for the acceleration of the second vehicle. ing. For this reason, “i _j ” is multiplied by 2 because even-numbered elements represent acceleration.

  Since the behavior of an automatically controlled vehicle is relatively easy to model, for example, passive dynamics f (x (t)) may be defined in advance as shown in the following equation (B).

(B)

  In the above formula (B), the speed and acceleration of each vehicle are arranged in order in [f (x (t))] as in the above formula (A), and the odd-numbered elements represent the speed. The relationship is as shown in the above formula (B).

As described above, the state space and the state space equation are defined according to various N, M, i ₁ , i ₂ ,..., I _M and the target steady state.

  The learning data collection unit 118 collects the state variable of the manually controlled vehicle, the state variable of the automatically controlled vehicle, the state variable of the automatically controlled vehicle, which are collected in advance for each state space representing the state of a plurality of vehicles including the manually controlled vehicle and the automatically controlled vehicle. A plurality of pieces of learning data representing the amount of change in the state variable and the amount of change in the state variable of the automatically controlled vehicle are collected.

Specifically, the learning data collection unit 118 collects learning data {(x ⁽ⁱ⁾ , f ⁽ⁱ⁾ )} ^D _{i = 1} corresponding to the state space for each state space. In this embodiment, since passive dynamics f (x (t)) R ^2N is modeled by a Gaussian process, D sets of learning data satisfying f (x ⁽ⁱ⁾ ) = f ⁽ⁱ⁾ {(x ⁽ⁱ⁾ , F ⁽ⁱ⁾ )} ^D _{i = 1} is prepared in advance.

As a method of collecting learning data {(x ⁽ⁱ⁾ , f ⁽ⁱ⁾ )} ^D _{i = 1} , for example, N vehicle groups at a certain time are picked up from past traffic data and {(x ^{( i)} , f ⁽ⁱ⁾ )} ^D _{i = 1} . In this case, for example, learning data {(x ⁽ⁱ⁾ , f ⁽ⁱ⁾ )} ^D _i including state variables and passive dynamics for each of a plurality of vehicles from an image acquired by a fixed point camera (not shown). _{= 1} is acquired.

  Note that the learning data collection unit 118 uses, as learning data, a gaussian for each vehicle using a vehicle state variable, vehicle state variables around the vehicle, and a change amount of the vehicle state variable. A process model may be learned.

  In this case, for example, it is assumed that the interaction between the N vehicles is only a nearby vehicle, and the approximation shown in the following equation (3) can be used.

(3)

In order to approximately decompose N vehicle groups as shown in the above equation (3), a vehicle group representing two vehicles or three vehicles as shown in the following equation (4) is used. Use multiple picked up data. As a result, the learning data collection unit 118 generates learning data (x ⁽ⁱ⁾ , f ⁽ⁱ⁾ ) for the N vehicle groups.

(4)

  The learning database 120 stores a plurality of learning data collected by the learning data collection unit 118.

  The state space model learning unit 122 is a model for predicting changes in state variables of a plurality of vehicles including a manually controlled vehicle and an automatically controlled vehicle based on a plurality of learning data stored in the learning database 120. Learn Gaussian process model. According to the Gaussian process model, the state variables of the plurality of vehicles are input, and the average of the amount of change of the state variables of the plurality of vehicles and the variance of the amount of change of the state variables of the plurality of vehicles are obtained.

Specifically, the state space model learning unit 122 uses the D sets of data {(x ⁽ⁱ⁾ , f ⁽ⁱ⁾ )} ^D _{i = 1} stored in the learning database 120 to use the passive dynamics f (x ).

  For example, the state space model learning unit 122 models the passive dynamics f (x) using a Gaussian process model based on the technique described in Reference Document 1.

Reference 1: C. E. Rasmussen and C. K. I. Williams, "Gaussian Processes for Machine Learning", MIT Press, 2014.

  The characteristics of the Gaussian process are as follows: (1) No knowledge of the modeling target is required and the prediction performance is high, (2) Overlearning is less likely to occur, (3) Not only the expected value of prediction but also the variance is obtained. Certainty can also be obtained as information.

Note that the following notation is used for learning data {(x ⁽ⁱ⁾ , f ⁽ⁱ⁾ )} ^D _{i = 1} .

(5)

(6)

  The state space model learning unit 122 models the passive dynamics f (x) as a Gaussian process using the learning data (X, F) represented by the above equations (5) and (6). When it is assumed that the passive dynamics f (x) follows a Gaussian process, the following equation (7) is established.

(7)

Here, _{K XX} ∈R ^{D × D} in the formula _{(7), K Xx (x} ) ∈R D, K xx (x) ∈R the kernel function ^{K (x (i), x} (j)) elements Are defined by the following equations (8) to (10).

(8)

(9)

(10)

In the above equation (8), δ _ij is the Kronecker delta, and σ _n εR is a hyperparameter. Also, _K xx in _K Xx in the above formula (9) (x) and the formula (10) (x) is a function of x. There are various types of kernel functions K (x ⁽ⁱ⁾ , x ^(j) ), but here, a Gaussian kernel represented by the following equation (11) is adopted.

(11)

Here, σ _f ∈R, Σ _w > 0∈R ^{n × n} in the above equation (11) is a hyperparameter, and Σ _w is a diagonal matrix. The state space model learning unit 122 performs hyperparameter θ: = {σ _n so that the log likelihood of [F] _{·, k} when the state X is given is maximized according to the following equation (12). , Σ _f , Σ _w }.

(12)

  However, since the passive dynamics f (x) modeled in the Gaussian process is n-dimensional, it is necessary to extend the framework of maximum likelihood estimation for the scalar function. In this embodiment, it is assumed that each component of the passive dynamics f (x) is independent, and the hyperparameter is expressed by the maximization problem shown in the following equation (13) so that all components share the same hyperparameter θ. Estimate θ.

(13)

  After the hyperparameter θ is estimated by the above equation (13), the predicted value f ^ (x) of the passive dynamics f (x) for a certain state x is an average E expressed by the following equations (14) to (16). It is given by a Gaussian distribution with [f ^ (x)] and variance Var [f ^ (x)].

(14)

(15)

(16)

  Here, in the present embodiment, it is assumed that the variance is sufficiently small to be negligible as shown in the following formula (C).

(C)

  Thus, the passive dynamics f (x) is modeled as the predicted average value E [f ^ (x)] of the Gaussian process.

In the present embodiment has been modeled together f (x), after the model separately for each f _{i ~} of the formula (3) ^(x), f _i according to the equation (3) ^˜ (x) may be combined to obtain f ^ (x).

  The control law design unit 124 controls the state quantity of the automatically controlled vehicle among the plurality of vehicles based on the passive dynamics learned for the state space by the state space model learning unit 122 for each state space. Is generated.

The control law design unit 124 calculates the state feedback control law u _FB (x) with respect to the average state variable variation matrix represented by the passive dynamics f ^ (x) obtained by the state space model learning unit 122 and the matrix B. To design. For example, the control law design unit 124 may use off-line model predictive control based on the technique described in Reference 2.

Reference 2: A. Grancharova, J. Kocijan and T. A. Johansen, "Explicit Stochastic Nonlinear Predictive Control Based on Gaussian Process Models.", ECC, 2007.

  Further, the control law design unit 124 may generate a control law based on the technique described in Reference Document 3.

Reference 3: R. W. Beard, G. Saridis and J. Wen, "Approximate Solutions to the Time-Invariant Hamilton-Jacobi-Bellman Equation", JOURNAL OF OPTIMIZATION THEORY AND APPLICATION, 1998.

  When the control law design unit 124 generates a control law based on the technique described in Reference Document 3, the residual of the GHJB equation is reduced using a generalized Hamilton Jacobi Bellman (GHJB) equation and a certain basis function. Thus, the control law is also obtained by approximately obtaining the solution of GHJB.

  Therefore, in this embodiment, the control law design unit 124 uses a predetermined basis function as a control law that optimizes the passive dynamics learned by the state space model learning unit 122 and the evaluation function expressed using the control law. A control law is generated by obtaining according to the HJB equation using. Examples of the basis function include a power polynomial of a state variable, a Chebyshev polynomial of a state variable, or a neural network that receives a state variable as an input. This will be specifically described below.

The design method of u _FB (x) using the HJB equation is shown below. Equation (17) below represents a continuous-time nonlinear system that is affine to the input. Further, consider a nonlinear optimal control problem for the evaluation function J (u, x (0)) as shown in the following equations (18) and (19). The evaluation function J (u, x (0)) is an example of an evaluation function expressed using a control law.

(17)

(18)

(19)

Here, q (x)> 0∈R in the above equation (18) represents a parameter for designing the cost of each time with respect to the state x, and R (x)> 0∈R ^{m × m} represents each of the inputs u. Represents a parameter for designing the cost of time. q (x) and R (x) are preset by the designer.

  In this case, the purpose is to derive the optimum control input u (x) that minimizes the evaluation function J (u, x (0)) as shown in the above equation (19). It is known that the nonlinear optimal control law that minimizes the above equation (18) for a continuous time nonlinear system that is affine to the input shown in (17) follows the HJB equation shown in the following equation (20). Yes.

(20)

  The solution V (x) of the above equation (20) is called a value function and needs to satisfy the following equations (21) to (24).

(21)

(22)

(23)

(24)

  When the value function V (x) satisfying the above equation (20) and the above equations (21), (22), (23), (24) is obtained, the optimum control input u (x) of the above equation (19) is It is given by the following equation (25).

(25)

However, it is generally difficult to solve the above equation (20), and there are many cases where the value function V (x) is not an elementary function. Therefore, in this embodiment, the candidate V (x, v ^~) value function is approximated by a basis function phi (x) as shown in formula (26).

(26)

Here, ^{v ~} ∈R ^Nb in the formula (26) is a coefficient for basis functions φ (x) ∈R ^Nb. As the basis function, there are a power polynomial of a state variable, a Chebyshev polynomial of the state variable, and a neural network having the state variable as an input. Subsequently, with respect to HJB equation of the equation (20), as shown in the following equation (27), the respective states X∈S _x residuals _r LSQ the square error obtained by integrating the (V ^(v ~)) Define.

(27)

Here, S _x in the above equation (27) is a control region relating to _{x, and} is set in advance by the designer. Then, as shown in the following equation (28) to minimize with respect to the ^{v ~} residuals ^{_{r LSQ (V (v ~)}} ).

(28)

By solving the optimization problem of the above formula (28) while updating the v ^~ a gradient method or the like appropriate initial values v ^{~ {0}} in origin, V (x, v ^~) is obtained with a high accuracy of approximation . When the finally obtained parameters are v ^{to *} , the control law u _FB (x) is given by the following equation (29) according to the above equation (25).

(29)

The control law design unit 124 stores the control law u _FB (x) obtained by the above equation (29) in the control law database 126.

The control law database 126 stores the control law u _FB (x) for each state space obtained by the control law design unit 124. The state space varies depending on N, M, i ₁ , i ₂ ,..., I _M. Further, a target steady state is set in advance according to the state space.

The control unit 128 reads the control law u _FB (x) stored in the control law database 126 in accordance with a control signal transmitted from the vehicle control amount calculation device 14 described later, and the vehicle control amount via the communication unit 130. It transmits to the calculation device 14.

The communication unit 130 transmits the control law u _FB (x) read by the control unit 128 to the vehicle control amount calculation device 14.

[Vehicle control amount calculation device]

  The vehicle control amount calculation device 14 includes a CPU that controls the vehicle control amount calculation device 14 as a whole, a ROM as a storage medium that stores programs of each processing routine, a RAM that temporarily stores data as a work area, and a connection between them. It is comprised by the computer containing the bus which carries out. In the case of such a configuration, a program for realizing the function of each component is stored in a storage medium such as a ROM or HDD, and each function is realized by executing the program by the CPU. To.

  When this computer is described in terms of functional blocks divided for each function realizing means determined based on hardware and software, the vehicle control amount calculating device 14 includes a state acquisition unit 140, a control law, as shown in FIG. The acquisition unit 142, the communication unit 144, the control law storage unit 146, and the control amount calculation unit 148 are provided.

  The state acquisition unit 140 acquires a state variable for each of a plurality of vehicles and information indicating which of the manually controlled vehicle and the automatically controlled vehicle is in a target traffic environment. Further, the state acquisition unit 140 determines a state space to which a plurality of vehicles belong based on the number of vehicles, the number of automatically controlled vehicles, and information on whether each vehicle is a manually controlled vehicle or an automatically controlled vehicle. Identify.

Specifically, the state acquisition unit 140 sequentially detects the positions and speeds of a plurality of vehicles in a target traffic environment. As a method for detecting the position and speed of a plurality of vehicles, for example, the position and speed of the plurality of vehicles in a target traffic environment can be detected from an image captured by a fixed point camera (not shown). Alternatively, the state acquisition unit 140 collects the positions and speeds of a plurality of vehicles acquired using inter-vehicle communication. Then, the state acquisition unit 140 selects appropriate N, M, i ₁ , i ₂ ,..., I _{M based on} the obtained position of the plurality of vehicles and information on whether the vehicle is a manually controlled vehicle or an automatically controlled vehicle. Is determined, and the current state variable x (t) is obtained.

  The control law acquisition unit 142 acquires the control law generated by the control law generation device 12 according to the state space specified by the state acquisition unit 140. Specifically, the control law acquisition unit 142 transmits a control signal including the state space specified by the state acquisition unit 140 to the control law generation device 12 via the communication unit 144.

The communication unit 130 of the control law generation device 12 receives the control signal transmitted from the vehicle control amount calculation device 14. The control unit 128 of the control law generation device 12 reads the control law u _FB (x) stored in the control law database 126 according to information representing the state space included in the control signal received by the communication unit 130. Specifically, the control unit 128 reads the control law u _FB (x) corresponding to the state space included in the control signal from the control law database 126. Then, the control unit 128 transmits the control law u _FB (x) to the vehicle control amount calculation device 14 via the communication unit 130.

The control law acquisition unit 142 acquires the control law u _FB (x) transmitted from the control law generation device 12 and stores the control law u _FB (x) in the control law storage unit 146. Further, the control law acquisition unit 142 acquires a target steady state from the control law generation device 12 according to the state space specified by the state acquisition unit 140.

The control law storage unit 146 stores the control law u _FB (x) acquired by the control law acquisition unit 142. The control law u _FB (x) stored in the control law storage unit 146 is a control law corresponding to the current state space.

Based on the state variables of the plurality of vehicles acquired by the state acquisition unit 140 and the control law u _FB (x) stored in the control law storage unit 146, the control amount calculation unit 148 The control amount of the state variable for the automatically controlled vehicle is calculated. Specifically, the control amount calculation unit 148 uses the control law corresponding to the current state space stored in the control law storage unit 146 for the state variables x (t) of the plurality of vehicles acquired by the state acquisition unit 140. Input to u _FB (x), and the control variable u (t) = u _FB (x (t))) of the state variable at the current time is calculated. Then, the control amount calculation unit 148 calculates and outputs the acceleration from the control amount u (t) of the state variable for the automatically controlled vehicle at the current time and the target steady state.

  The acceleration output by the control amount calculation unit 148 is given to a plurality of automatic control vehicles.

<Operation of Control Law Generating Device 12>
Next, the operation of the control law generation device 12 according to the present embodiment will be described. First, a plurality of learning data is collected by the learning data collection unit 118 and stored in the learning database 120. And the state space model learning part 122 performs the learning process routine shown in FIG.

  In step S100, the state space model learning unit 122 acquires a plurality of pieces of learning data of the state space stored in the learning database 120 for each state space.

  In step S102, the state space model learning unit 122 predicts the amount of change in the state variables of the plurality of vehicles based on the plurality of learning data of the state space acquired in step S100 for each state space. As a Gaussian process model.

  In step S104, the state space model learning unit 122 outputs the Gaussian process model learned for each state space as a result, and ends the learning processing routine.

  Next, the control law design unit 124 generates a control law for an automatically controlled vehicle among a plurality of vehicles based on the Gaussian process model learned for each state space by the learning processing routine. When the learning process routine ends, the control law design unit 124 executes the control law design process routine shown in FIG.

  In step S200, the control law design unit 124 acquires the average of the Gaussian process models learned for each state space by the learning processing routine and the coefficient matrix B for each state space.

In step S202, the control law design unit 124 calculates the control law u _FB (x) for each state space according to the above equation (29) based on the average of the Gaussian process model and the coefficient matrix B acquired in step S200. To get to.

In step S204, the control law design unit 124 stores the control law u _FB (x) for each state space obtained in step S202 in the control law database 126, and ends the control law design processing routine.

<Operation of Vehicle Control Amount Calculation Device 14>
Next, the operation of the vehicle control amount calculation device 14 according to the present embodiment will be described. The vehicle control amount calculation device 14 executes a control amount calculation processing routine shown in FIG. 4 under the target traffic environment.

  In step S300, the state acquisition unit 140 acquires a state variable x (t) for each of a plurality of vehicles and information on which of the manually controlled vehicle and the automatically controlled vehicle is in a target traffic environment. To do. Further, the state acquisition unit 140 determines a state space to which a plurality of vehicles belong based on the number of vehicles, the number of automatically controlled vehicles, and information on whether each vehicle is a manually controlled vehicle or an automatically controlled vehicle. Identify.

In step S302, the control law acquisition unit 142 acquires the control law stored in the control law database 126 of the control law generation device 12 according to the state space specified in step S300. Then, the control law acquisition unit 142 stores the acquired control law u _FB (x) in the control law storage unit 146. Further, the control law acquisition unit 142 acquires a target steady state from the control law generation device 12 according to the specified state space.

In step S304, the control amount calculation unit 148 includes the plurality of vehicle state variables x (t) acquired by the state acquisition unit 140, and the control law u _FB (x stored in the control law storage unit 146 in step S302 above. ), A control variable u (t) of a state variable for an automatically controlled vehicle among a plurality of vehicles is calculated. Then, the control amount calculation unit 148 calculates and outputs the acceleration for each of the automatic control vehicles from the control amount u (t) of the state variable for the automatic control vehicle at the current time and the target steady state.

  In step S306, the control amount calculation unit 148 outputs the acceleration for each of the automatically controlled vehicles calculated in step S304 as a result, and ends the control amount calculation processing routine.

  As described above, according to the control law generation device of the present embodiment, for each state space representing a plurality of vehicle states, a plurality of state variables for a manually controlled vehicle and a state variable for an automatically controlled vehicle collected in advance. Based on the learning data, a model for predicting the amount of change in the state variables of a plurality of vehicles is learned, and for each state space, an automatically controlled vehicle among the plurality of vehicles based on the model learned for the state space By generating the control law, it is possible to obtain a control law for controlling the automatic control vehicle in a traffic environment including the manual control vehicle and the automatic control vehicle.

  In addition, according to the control amount calculation device of the present embodiment, the state variable for each of the plurality of vehicles and the information on which of the manual control vehicle and the automatic control vehicle are acquired, and the plurality of vehicles are Identify the state space to which it belongs, acquire the control law generated by the control law generation device according to the specified state space, and based on the state variables of the plurality of vehicles and the acquired control law, By calculating the control amount for the automatic control vehicle among the vehicles, it is possible to obtain the control amount for controlling the automatic control vehicle even in a traffic environment including the manual control vehicle and the automatic control vehicle.

  In addition, in a traffic environment including a manually controlled vehicle in which behavior is difficult to predict, it is possible to execute a safe and optimal control of the automatically controlled vehicle in cooperation with the manually controlled vehicle in real time. This is expected to reduce traffic congestion and improve fuel efficiency and travel time.

  In addition, from the learning data including the behavior data of manually controlled vehicles, multiple vehicles with manually controlled vehicles are regarded as one system and modeled in a Gaussian process, so that the prediction of the entire system including manually controlled vehicles is highly accurate. Can be realized. In addition, the design in which a plurality of vehicles are regarded as one model enables vehicle control in cooperation with a manually controlled vehicle whose behavior is difficult to predict.

  Also, by designing the control law off-line based on the learned model, it is only necessary to calculate the control input in accordance with the control law online, so that control with high real-time characteristics is possible.

  In the above embodiment, the control in the front-rear direction of the automatically controlled vehicle has been described as an example. However, the present invention is not limited to this, and control including the width direction (lateral direction) of the vehicle may be performed.

  The program of the present invention can be provided by being stored in a recording medium.

DESCRIPTION OF SYMBOLS 12 Control law production | generation apparatus 10 Vehicle control system 14 Vehicle control amount calculation apparatus 118 Learning data collection part 120 Learning database 122 State space model learning part 124 Control law design part 126 Control law database 128 Control part 130 Communication part 140 State acquisition part 142 Control Law acquisition unit 144 Communication unit 146 Control law storage unit 148 Control amount calculation unit

Claims (7)

The state of the manually controlled moving body collected in advance for each state space representing the state of a plurality of moving bodies including a manually controlled moving body that is a manually controlled moving body and an automatically controlled moving body that is a automatically controlled moving body. A learning unit that learns a model for predicting the amount of change in the state variables of the plurality of moving bodies based on a plurality of learning data representing the variables and the state variables of the automatically controlled moving bodies;
For each state space, based on the model learned for the state space by the learning unit, a control law is generated for controlling each state variable of the automatic control moving body among the plurality of moving bodies. A control law design department to
Model learning device including The learning unit, as the learning data, for each of the moving bodies, the state variable of the moving body, the state variable of the moving body around the moving body, the amount of change of the state variable of the moving body, The model learning apparatus according to claim 1, wherein the model is learned by using a model. The control law design unit includes the model learned by the learning unit and the control law for optimizing an evaluation function represented by using the control law, a power polynomial of the state variable, Chebyshev of the state variable The model learning apparatus according to claim 1, wherein the control law is generated by obtaining a polynomial or a neural network having the state variable as an input according to an HJB equation having a basis function. The learning unit receives a plurality of state variables of the mobile objects, and outputs an average of change amounts of the state variables of the mobile objects and a variance of change amounts of the state variables of the mobile objects. The model learning apparatus according to any one of claims 1 to 3, wherein a process model is learned as the model. The state variable includes the position and speed of the moving body,
The model learning device according to any one of claims 1 to 4. A state acquisition unit for acquiring a state variable for each of the plurality of moving bodies and information on whether the moving body is a manual control moving body or an automatic control moving body, and identifying a state space to which the plurality of moving bodies belong; ,
According to the state space specified by the state acquisition unit, a control law acquisition unit that acquires the control law generated by the model learning device according to any one of claims 1 to 5,
A control amount for calculating a control amount of the state variable for the automatic control mobile body among the plurality of mobile bodies based on the state variables of the plurality of mobile bodies and the control law acquired by the control law acquisition unit A calculation unit;
Control amount calculation apparatus including Computer
The state of the manually controlled moving body collected in advance for each state space representing the state of a plurality of moving bodies including a manually controlled moving body that is a manually controlled moving body and an automatically controlled moving body that is a automatically controlled moving body. Based on a plurality of learning data representing a variable and a state variable of the automatic control moving body, a learning unit for learning a model for predicting a change amount of the state variable of the plurality of moving bodies, and for each state space, Based on the model learned for the state space by the learning unit, a control law design unit that generates a control law for controlling each state variable of the automatic control mobile body among the plurality of mobile bodies A program to make it work.