Control device, traffic rectification system, control method, and program

Abstract

A control device provided in each of a plurality of moving bodies that are provided in a traffic rectification system wherein the plurality of moving bodies autonomously perform traffic rectification to prevent collisions between these moving bodies, said control device being provided with: a state update unit that updates the states of the moving bodies under constraints for preventing collisions between the moving bodies, on the basis of state update dynamics including sub-dynamics for updating the states of the moving bodies, and sub-dynamics for message passing between each moving body and other moving bodies in proximity; and an output unit that outputs the states updated by the state update unit and a message.

Description

[Technical Field]

[0001] The present invention relates to a technology for autonomously rectifying traffic (which may also be called traffic control) using a plurality of vehicles. [Background technology]

[0002] In conventional technology, traffic lights installed at road intersections and other locations are used to regulate traffic so that vehicles can travel safely without crashing. However, this conventional technology has the problem of causing chronic congestion in urban areas due to frequent stops while waiting for traffic lights to change and temporary stops when merging or turning right or left. [Prior art documents] [Non-patent literature]

[0003] [Non-Patent Document 1] NTT, "IOWN: Mobility by IOWN (concept video)", https: / / www.youtube.com / watch?v=4fo_kEYrY6E Summary of the Invention [Problem to be solved by the invention]

[0004] As autonomous vehicles become more widespread in the future, it is expected that the centralized control systems using conventional traffic lights will be replaced with distributed control systems (signal-free traffic rectification), and that the above issues will be resolved using distributed control systems.

[0005] The present invention has been made in consideration of the above points, and aims to provide a technology for realizing traffic control without the use of traffic lights by having each vehicle perform autonomous control. [Means for solving the problem]

[0006] According to the disclosed technology, there is provided a traffic control system including a plurality of mobile bodies each having a control device, the plurality of mobile bodies autonomously performing traffic control so as to prevent collisions between the mobile bodies, the control device comprising: a state update unit that updates the state of the moving object based on state update dynamics having sub-dynamics for updating the state of the moving object and sub-dynamics for message passing between the moving object and other moving objects in the vicinity of the moving object under constraints for preventing collisions between the moving objects; an output unit that outputs the state updated by the state update unit and a message; the state update unit executes a process of alternately repeating the state update of the moving object and the message passing by using a neural network model for solving an initial value problem of an ordinary differential equation that expresses the state update dynamics; The parameters in the neural network are updated by backpropagation calculation based on the adjoint method so that the velocity of the moving object approaches the target velocity under the constraints. A control device is provided. [Effects of the Invention]

[0007] The disclosed technology provides a technology for realizing traffic regulation without the use of traffic lights by having each vehicle autonomously control the traffic. [Brief explanation of the drawings]

[0008] [Figure 1] FIG. 1 is a diagram illustrating an example of a system configuration according to an embodiment. [Figure 2] FIG. 1 is a diagram for explaining an overview of an embodiment. [Figure 3] FIG. 1 is a diagram for explaining an overview of an embodiment. [Figure 4] FIG. 1 illustrates an example of a system configuration. [Figure 5] FIG. 1 illustrates a NODE-based DNN architecture. [Figure 6] FIG. 1 illustrates an extended NODE-based DNN architecture that follows a state-space model. [Figure 7] FIG. 1 illustrates a NOS-based DNN architecture. [Figure 8] FIG. 1 illustrates Algorithm 1. [Figure 9] FIG. 10 is a diagram showing representative parameters. [Figure 10] FIG. 10 is a diagram for explaining an experiment. [Figure 11] FIG. 10 is a diagram for explaining an experiment. [Figure 12] FIG. 10 is a diagram for explaining an experiment. [Figure 13] FIG. 10 is a diagram for explaining an experiment. [Figure 14] FIG. 10 is a diagram showing experimental results. [Figure 15] FIG. 10 is a diagram showing experimental results. [Figure 16] FIG. 10 is a diagram showing experimental results. [Figure 17] FIG. 2 is a diagram illustrating an example of a functional configuration of a vehicle 1. [Figure 18] FIG. 2 is a diagram illustrating a functional configuration of a control device 100. [Figure 19] FIG. 2 is a diagram illustrating the functional configuration of a control server 200. [Figure 20] FIG. 2 illustrates an example of a hardware configuration of the apparatus. DETAILED DESCRIPTION OF THE INVENTION

[0009] Hereinafter, an embodiment of the present invention (the present embodiment) will be described with reference to the drawings. The embodiment described below is merely an example, and the embodiment to which the present invention is applied is not limited to the following embodiment.

[0010] (Outline of the embodiment) Figure 1 shows an example of the configuration of a traffic rectification system according to this embodiment. As shown in Figure 1, this traffic rectification system has multiple nodes, and each node communicates wirelessly with other nearby nodes. The connections between nodes (connections for wireless communication) are called edges. In this traffic rectification system, for example, vehicles can move without colliding with each other.

[0011] Although the node is not limited to a specific object, in this embodiment, it is assumed that the node is a vehicle traveling on a road. Hereinafter, the node may be called a "vehicle." The node may also be called a "mobile body."

[0012] FIG. 2 shows an image of multiple vehicles traveling on a road. Each vehicle autonomously controls its speed based on communication with nearby vehicles and state updates using a DNN (deep neural network) to avoid collisions with other vehicles and to approach a target speed. This is expected to reduce travel and transportation times to the maximum extent possible while preventing traffic accidents (collisions). Note that in this embodiment, only speed is used as a state, but this is an example. It is also possible to use something other than speed, or both speed and something other than speed, as a state. States other than speed include, for example, route, lane, steering direction, etc.

[0013] The neural network model installed in each vehicle is a model for solving an initial value problem of an ordinary differential equation that represents state update dynamics. The ordinary differential equation can be expressed, for example, as follows:

[0014] dx / dt=M1(x,t,θ,A,b)+M2(x,t,A,b) As mentioned above, the state update dynamics is decomposed into M1 and M2, where M1 corresponds to intra-vehicle state updates and M2 corresponds to close-proximity inter-vehicle communication (message passing).

[0015] For each vehicle, the state (x) is updated according to the dynamics (M1+M2) under the collision prevention constraint (Ax+b≦0). Detailed examples of M1 and M2 will be described later.

[0016] In this embodiment, ordinary differential equations are discretized to form a control algorithm (neural network) that alternately repeats (1) updating of the in-vehicle state and (2) inter-vehicle communication.

[0017] Figure 3 shows an image of the above control algorithm when the number of vehicles is N. As shown in Figure 3, after initialization, updates to the in-vehicle state within each vehicle and communications between each vehicle and its neighboring vehicles are repeated alternately. As shown in Figure 3, the overall calculation is distributed and parallelized, so the calculation load on each vehicle is light. In addition, for inter-vehicle communications, proximity sparse communication is used to avoid straining the network bandwidth.

[0018] The neural network that performs state update as described above uses the adjoint method for learning (learning the parameter θ). That is, an efficient differential calculation is performed using backpropagation calculations of the neural network based on the adjoint method to learn the parameter (θ) so that the average speed accelerates toward the target speed under the constraint of collision prevention. In this embodiment, the parameter (θ) is common to all vehicles, but it is also possible to learn so that each vehicle uses a different parameter θ.

[0019] FIG. 4 is a diagram showing a system in more detail when the nodes are vehicles. As shown in FIG. 4, there are multiple vehicles 1, and each vehicle is equipped with a control device 100. The control device 100 is equipped with the above-mentioned DNN model and performs state acquisition, state update, and state output. In the example of FIG. 4, a control server 200 is provided. Here, the control server 200 receives information such as state from each vehicle (which may be a cost calculated by each vehicle), learns a parameter (θ), and transmits the learned parameter θ to each vehicle.

[0020] Incidentally, the provision of the control server 200 is just an example. The control server 200 may not be provided, and each vehicle may learn the parameter θ by itself.

[0021] Below, the ordinary differential equations mentioned above and specific examples of techniques for solving the initial value problems thereof will be described in detail.

[0022] (NOS Overview) Neural Ordinary Differential Equation (NODE) ​​has been proposed as a DNN model for solving initial value problems (IVP) of ordinary differential equations.

[0023] In NODE, an IVP is formulated by giving a state variable x(t) with an initial state x(0), a learnable parameter θ, and a nonlinear dynamics M with time t. Figure 5 shows the DNN architecture based on NODE. As shown in Figure 5, the states are updated sequentially and cost calculations are performed based on each state.

[0024] On the other hand, in this embodiment, an external control input o(t) is further introduced, which is not present in the NODE. In this case, the IVP can be defined by the following equation (1).

[0025] ∂x / ∂t=M(x,o,θ,t) (1) Note that the time indexes x and o in the above equation (1) are omitted for simplicity of notation.

[0026] Since NODE-based DNN architectures can be constructed as discretizations of basic IVPs, various discretization methods in ODE solvers, such as high-order Runge-Kutta solvers, are available to construct various NODE-based DNN architectures.

[0027] Although NODE can handle basic IVPs, it is difficult to apply it to large-scale systems with complex dynamics M, which can be represented as a graph consisting of many subsystems (nodes) and their connections (edges). An example of such a system is the aforementioned system consisting of multiple vehicles that perform traffic control.

[0028] Furthermore, it is difficult for centralized computational methods such as NODE to perform the forward propagation (x transition) and backward propagation (θ learning) in equation (1) for large-scale systems.

[0029] In the system consisting of the above nodes and edges, it is effective to execute processing in a distributed manner by repeating (i) state transitions at each node (e.g., updating the speed state at each vehicle) and (ii) message passing between nodes (e.g., communication between vehicles located close to each other).

[0030] To address this issue, this embodiment employs federated dynamics learning using neural operator splitting (NOS) as an extension of NODE, as outlined below (details will be provided later).

[0031] The overall dynamics M is decomposed into (i) M1 for the state transition of each node and (ii) M2 for message passing between nodes to efficiently manage a large-scale system in graph form, and is expressed as shown in equation (2) below.

[0032] ∂x / ∂t=M1(x,o,θ,t)+M2(x,o,t) (2) Here, M1 can be further decomposed into sub-dynamics at each node. M2 is related to message passing and is therefore set to have no learnable parameters.

[0033] Furthermore, by discretizing the above (2) using operator decomposition methods such as the well-known techniques PRS (Peaceman-Rachford Splitting) and DRS (Douglas-Rachford Splitting), the state transitions can be in residual, recurrent, and alternating forms.

[0034] In addition, in this embodiment, state domain relaxation is performed to impose physical constraints. That is, in equation (2), the state domain of M1 and the state domain of M2 are the same as x, but in the NOS of this embodiment, the constraints on the state domain of equation (2) are relaxed to impose physical constraints on the state variables of many nodes.

[0035] Specifically, x is used for M1, and {A, b} is given for M2, and the auxiliary state variable y = Ax + b is used to set M2(y, o, t). y = Ax + b represents a physical constraint, for example, y = Ax + b ≦ 0. That is, (i) the speed of each vehicle (N nodes {x1,...,x N}∈x) is updated according to M1, while (ii) a certain distance from other vehicles is maintained by designing M2 as y=Ax+b≦0, for example, using an inequality. This appropriately limits the parameter search area, and is expected to prevent vehicle collisions and enable rapid and stable learning of θ.

[0036] As will be described later, in this embodiment, the NOS is associated with a constrained cost minimization problem derived by the ADMM (Alternating Direction Method of Multipliers) method.

[0037] (Extended NODE according to the state space model) Before providing a detailed explanation of NOS, we will explain the extended NODE based on the state-space model as a related technology to NOS. This is an extended NODE based DNN architecture based on the state-space model that assumes a noisy, nonlinear, and indirect observation process for the state variable x in real applications.

[0038] In the following, we first define a cost function that takes into account the observation process. Next, we formulate the adjoint method for learning x and the parameter θ in the cost function using 4D-Var, and derive an extended NODE-based DNN architecture that follows the state space model.

[0039] <Cost function derivation> Assume that the state variable x(t) is updated sequentially at time intervals t∈(0,T) according to equation (1). Given the initial state x(0) and the external control input o(t), equation (1) can be rewritten as the following integral form, equation (3):

[0040]

number

[0041]

number

[0042] In equation (4), R(t) is the covariance matrix of the measurement noise ξ, T represents the transpose of a matrix. The measurement noise is Gaussian distributed ξ~Norm(0,υ 2 ), we can choose a scaled identity matrix as a particular R(t). Since the state transition from x(0) to x(T) is constrained by equation (1), the final cost form J(x,θ) for learning θ is given by the constrained cost integral minimization problem, as shown in equation (5) below.

[0043]

number

[0044]

number

[0045]

number

[0046] Furthermore, from equation (4),

[0047]

number

[0048] Initialized as λ(T)=0, the adjoint variable λ(t) is updated backward, resulting in λ(0)=g=[g T x0 ,g T θ ] T Here, g is the gradient of the cost function J at time 0. The gradient of the cost function can be obtained by backward integrating the adjoint equation with respect to time, as shown in equation (7) below.

[0049]

number

[0050] <NODEベースのDNNアーキテクチャ> Some NODE-based DNN architectures can be obtained by discretizing equation (1). For simplicity of notation, x k at discrete time t k Let D be the estimated system state at (k=1,...,K), and let D approximate the residual state update via M at discrete time intervals Δt.

[0051] Given the state variables {θ,o}, x k+1 =D(x k ) and updated in a residual and recursive form. D is q D q-1 ...D1 is allowed to consist of a set of operators.

[0052] The Forward Euler method (8), Backward Euler method (9), and Crank-Nicolson method (second-order Runge-Kutta) shown below can be used to approximate differential equations with discrete state transition rules.

[0053] [Forward Euler Method]

[0054]

number

[0055]

number

[0056]

number

[0057]

number

[0058] The NODE-based DNN architecture can be used for K-iteration recursive stacks based on residual state updates ((8), (9), or (10)), as shown in Figures 5 and 6. Note that the forward Euler algorithm can be considered as a ResNet.

[0059] The adjoint method can be used to calculate the gradient in training θ. In this case, the NODE-based DNN architecture is extended to follow Figure 6 by discretizing the continuous backpropagation (7), as shown in Equation (11) below.

[0060]

number

[0061]

number

[0062] Here, we introduce a NODE-based DNN architecture that follows a state-space model. However, this approach becomes computationally intractable when the dynamics of the overall system become complex, such as in the case of autonomous traffic control by multiple vehicles.

[0063] Therefore, in this embodiment, we introduce a NOS-based DNN architecture for federated dynamics learning, and solve the traffic rectification problem by alternating between vehicle state updates (i) and message passing (ii) between vehicles.

[0064] (Neural Operator Decomposition (NOS)) Below, we describe in detail the NOS-based DNN architecture for federated dynamics learning. First, we explain the discretization of Equation (2) using operator decomposition methods such as PRS and DRS. Next, we explain how NOS is related to the constrained cost minimization problem for domain relaxation of state variables in {M1, M2}, which is important for managing state variables in large-scale distributed systems. Furthermore, we identify the shape of {M1, M2} for federated dynamics learning, which involves iterating node state transitions (i) and message passing between nodes (ii).

[0065] <Discretization by operator decomposition> We will explain the discretization of equation (2) using the operator decomposition method. NOS can use various operator decomposition methods such as PRS, DRS, and FBS (Forward-Backward Splitting).

[0066] In this embodiment, we select PRS and DRS because they are based on the Crank-Nicholson algorithm (10) with second-order accuracy and are expected to achieve accurate state transitions.

[0067] Discretizing equation (2) using PRS leads to a residual, recursive, and alternating DNN architecture, as shown in equation (12) below. [PRS-Net]

[0068]

number

[0069]

number

[0070] The state transition is a set of residual operators D=D q D q-1 Since the learning of θ is performed by D1, the gradient of the learning of θ is T =D1 T D2 T ···D q T can be obtained by replacing

[0071] <Relaxing the state domain for flexible subdynamics coupling> Here, we relate NOS (PRS-Net and DRS-Net) to a constrained cost minimization problem for state-domain relaxation of two sub-dynamics {M1, M2}.

[0072] When the state variable of M1 is x(t), the state variable of M2 is relaxed by y(t) = A(t)x(t) + b(t). Here, we omit the time index {A, b} to simplify the notation. Also, for any t, the inverse matrix (AT A) -1 Let A be chosen such that there exists a

[0073] For the ADMM flow, this takes into account the discrete-time restriction to create a (discrete) ADMM in continuous ODE form. In the ADMM flow, {M1,M2} are (A T A) -1 is specified by the (sub)derivative of the cost function V to be minimized under an affine transformation using

[0074] Based on this, in this embodiment, {M1, M2} is set by decomposing the function V as V = V1 + V2 as shown in the following equation (14), where V1 is a smooth but non-convex function, and V2 is a convex function but may not be differentiable.

[0075]

number

[0076] To investigate the physical meaning of (14), we relate it to a constrained cost minimization problem. For this purpose, we restrict {V1,V2} to be convex and the constraint parameters {A,b} to be time-invariant. The linearly constrained convex minimization problem is given by (15):

[0077]

number

[0078]

number

[0079] Applying DRS to equation (16) yields the well-known ADMM algorithm, which updates {x,y,z} in a residual, recursive, alternating fashion in discrete steps.

[0080] V1 in equation (14) may be non-convex, but when Δt is sufficiently small, the previous state x k-1 can be approximated by a locally quadratic form by Taylor expansion around

[0081]

number

[0082] <Cost function specification for learning associative dynamics> Here, for the purpose of associative dynamics learning, we define {V1, V2} in Equation (14). For this purpose, we introduce graph notation symbols. For the traffic rectification problem of this embodiment, the connections (edges) between vehicles (nodes) change over time. Therefore, we use a time-varying graph G(t).

[0083] G(t) consists of a set of N(t) nodes, N(t), and a set of E(t) edges, E(t). The set of indices of adjacent vehicles connected to the i-th node is E i (t) = {j∈N(t)|(i,j)∈E(t)}. Hereafter, we omit the time index {G(t),N(t),E(t)}.

[0084] Each node has its own cost function V local,i (x i ,o i ,θ i ,t). From a global perspective, the input variable stack is a time-varying variable x=[x1 T ,...,x N T ] T ,o=[o1 T ,...,o N T ] T and the time-invariant variable θ=[θ1 T ,...,θ N T ] T It consists of:

[0085] For simplicity of notation, the control input o i The edge information E i and state x i and the state of the adjacent node x j (j∈E i ) to relate the time-varying constraint parameters {A i|j ,b i|j} is included. i|j ,b i|j The stack of {A, b} is related to {A, b} in equation (14) as follows:

[0086]

number

[0087]

number

[0088] The inequality constraint can be expressed as (I+P)(Ax+b)≦0, where P(o,t) is a permutation matrix that specifies the exchange of variables between adjacent nodes according to the time-varying graph G. When the constraint is expressed as an indicator function, Equation (18) can be written as Equation (19) below.

[0089]

number

[0090]

number

[0091]

number

[0092] (Distributed Traffic Control for Autonomous Vehicles) We explain in detail the operation of a distributed control system that realizes traffic rectification using NOS. Each vehicle autonomously updates its state (e.g., speed) using external control inputs (e.g., inputs from cameras / LiDAR) and message passing between nearby vehicles.

[0093] The behavior of the distributed control system is a function of state transition dynamics, and the inputs are the previous state x and the control input o. Furthermore, we use NOS-based federated dynamics learning to optimize the parameter θ of the state transition dynamics.

[0094] Below, we first provide an overview of related technologies related to decentralized traffic control. Next, we explain {J, V, A, b} in equation (21) for signal-free traffic control, which accelerates the average vehicle speed to a target speed while maintaining inter-vehicle distance to avoid collisions. Finally, we explain the derivation of a state transition DNN architecture based on PRS-Net and DRS-Net.

[0095] <Related technologies> Related technologies related to traffic control include, for example, the technology disclosed in the reference "Wang, Z., Zheng, Y., Li, SE, You, K., and Li, K. (2018). Parallel optimal control for cooperative automation of large-scale connected vehicles via admm."

[0096] This technique uses ADMM to move each vehicle to a target speed while maintaining a certain distance between vehicles. However, it has the following problems: (i) extra link nodes are required to connect vehicle nodes; (ii) there are no learnable parameters in the state transition dynamics.

[0097] In contrast, the technology according to the present embodiment (i) realizes a completely decentralized traffic control consisting only of vehicle nodes, and (ii) realizes federated dynamics learning to obtain optimal autonomous state transition dynamics.

[0098] <Cooperative Dynamics Learning for Distributed Traffic Control> First, for the traffic flow problem, we use {J, V local,i} is formulated.

[0099] For N autonomous vehicles (nodes), their routes, initial speed states, positions, and maximum speeds s max Given [m / s], each vehicle has a normalized velocity state x=[x1,...,x N ] T ,(x i ∈[0,1]) ― θ(=θ1=,...,=θ N ) is considered to be retained. In addition, due to the restrictions on the character type of the specification, the bar at the beginning of the character is written as " ― It is written before the letter, such as "θ".

[0100] The goal of federated dynamics learning for decentralized signal-free traffic control is to learn how to move vehicles in a certain distance from each other while avoiding collisions. i (t) is the target x tar ∈[0,1] ― The goal is to find θ.

[0101] Therefore, in this embodiment, the following function is selected as the cost function in equation (4):

[0102]

number

[0103] Each vehicle drives in the center of a single-lane road and overtaking is prohibited, i.e., each vehicle follows the other vehicles across the intersection. i is the normalized velocity o spd,i , a 2D direction vector o dir,i , 2D position vector o pos,i , the surrounding image of each vehicle o img,i , mapping vector o map,i|j , and the set of adjacent connections j∈E i The mapping vector o map,i|j is used to convert the 2D position into a scalar value measuring the distance from the i-th vehicle to the j-th vehicle.

[0104] Cost function - V local (=V local,1 =,…,=V local,N ) is expressed as a quadratic function of the velocity state x i (t). The quadratic function is designed to estimate the acceleration

[0105]

number

[0106]

number

[0107] ― V localBy substituting into equation (12) for PRS-Net and into equation (13) for DRS-Net, the residual, recursive, and alternating update rules shown in Figure 7 can be summarized as the following equation (22).

[0108]

number

[0109]

number

[0110] Next, the determination of {A, b} in equation (18) will be described. In this embodiment, the state pair {x i ,x j} is subject to a non-collision constraint. i ,x j} is expressed as follows: min Manhattan distance (l i|j +l j|i ) is constrained to keep

[0111]

number

[0112]

number

[0113] From the update rule for x in equation (22), {A, b} is related to the acceleration / deceleration of each vehicle. For example, A i|j b i|j When is positive, the i-th vehicle brakes to maintain a certain distance from the j-th vehicle.

[0114] To avoid the problem of many vehicles being stuck because the rear vehicle cannot overtake the stopped vehicle in front, in this embodiment, a front flag / back flag is assigned to each vehicle pair, and the front vehicle is prevented from braking as much as possible. i|j ,b i|j The design of} is as shown in the following equation (24).

[0115]

number

[0116] We will now explain in more detail how to derive the above equation, i.e., the collision-free constraint in inequality form. i , 2D direction o dir,i , and the parameter {s max , Δt}, the position of the i-th vehicle at the next time step is p i =o pos,i +Δt s max x i ·o dir,i The position of the jth vehicle can be estimated as p j =o pos,j +Δt s max x j ·o dir,j It can be estimated as:

[0117] A pair of mapping vectors {o map,i|j ,o map,j|i} to set Manhattan distance

[0118]

number

[0119]

number

[0120] Next, we will explain how to convert equation (23) to equation (24). From the update rule for x in equation (22), A i|j b i|j We can see that vehicle i brakes when is positive. As mentioned above, {A i|j ,b i|j The design policy of} is to avoid collisions by assigning a front / rear flag to each vehicle pair and then adjusting the acceleration / deceleration of the rear vehicle.

[0121] In the following case, the front / rear assignment is determined by the position and direction of the vehicle pair. In the intersection crossing case, the front / rear flag assignment can be determined by the estimated remaining time to the intersection center using the current speed and position.

[0122] Assuming that the i-th vehicle is in front and the j-th vehicle is behind, the transformation to fit the inequality constraint in equation (23) to the inequality constraint in equation (18) is as shown in equation (24), where A i|j normalize to 1 and set all remaining biases to b j|i This is done by transferring the i|j ,A i|j} has been normalized, so A T All eigenvalues ​​of A are 1, i.e., σ 2 max =σ 2 min =1.

[0123] The procedure for distributed traffic rectification using PRS-Net and DRS-Net is outlined in Algorithm 1 (Alg.1) in Figure 8. In Algorithm 1, the operation of each node / edge is broken down and shown.

[0124] In the procedure of FIG. 8, initialization is performed in line 1. The processes in lines 3 to 19 are performed for each time k, where k is from 1 to K. In lines 4 to 6, each node i is assigned a control input, an edge connection, and a parameter {A i|j ,b i|j} is obtained.

[0125] In lines 8 to 11, each node i performs message passing (message z j|i (receive)

[0126] In lines 13 to 18, each node updates its internal state based on equation (22). As a result, node i updates the speed (x i ) and the k+1 message (z for each j i|j ) can be calculated.

[0127] In the above process, a single-dimensional {z i|j ,z j|i} is the number of connected vehicles (i, j∈E i ), so communication is lightweight.

[0128] Each vehicle (or the control server 200) records data {x, o, A, b} for a certain time window KΔt=T. Each vehicle (or the control server 200) performs back propagation (11) using the recorded data {x, o, A, b} to calculate - Update θ. Associative dynamics learning is performed by repeating I rounds of forward / backward propagation.

[0129] The example shown in FIG. 7 illustrates an image in which the control server 200 records data {x, o, A, b} at each discrete time and performs parameter learning of backward propagation based on cost calculation.

[0130] (Numerical experiment) To demonstrate the effectiveness of NOS-based federated dynamics learning, we conducted numerical experiments using the traffic simulator SUMO. We investigated the average vehicle speed while avoiding vehicle collisions.

[0131]

number

[0132] <Experimental setting> As shown in Figure 10, N = 30 vehicles were randomly placed in the traffic simulator. Ten road maps were also prepared. To avoid overfitting to the training data, each road map had small random perturbations in the length of straight roads and the locations of intersections.

[0133] The initial vehicle states (position and velocity) were randomly set. Each vehicle traveled on a randomly selected road predefined with right and left turns. For each mini-batch (with random road maps and road perturbations), we chose K = 300 iterations and Δt = 0.1 s.

[0134] Therefore, forward propagation through the traffic simulation is performed for T = KΔt = 30.0 seconds, and then the dynamics parameters are calculated using the recorded data {x, o, A, b}. ― Backpropagation was used to update θ. The parameters were g with a learning rate of 0.0025. θ This set of forward and backward propagation is one round.

[0135] Estimates of the dynamics parameters were obtained over a total of I = 500 rounds. Evaluation sessions were conducted every 10 rounds. To ensure a fair comparison between the proposed method and the reference method, a total of 50 fixed initial settings (10 road maps × 5 road randomness) were prepared.

[0136] Images of each vehicle and its surroundings are used as control inputs to determine the acceleration / deceleration of each vehicle. img,i was generated at each discrete time instant.

[0137] The image size is defined as 64 (W) × 64 (H) × 5 (Ch). Each channel consists of (1) the surrounding road, (2) the surrounding vehicle positions, (3) their normalized velocities, and (4, 5) the 2D direction vectors of the surrounding vehicles. Each image is rotated so that each vehicle faces the same direction. img Examples of these are shown in Figures 11 and 12, respectively. Figure 13 shows an image of repeating state updates and learning for each round.

[0138] In the proposed method (NOS), PRS-Net and DRS-Net are constructed based on Alg. 1. In the state update of Equation (22),

[0139]

number

[0140] To achieve realistic message passing, we assume that P is a sparse matrix. com = 25.0 m for a pair of vehicles (i, j∈E i ) can be communicated at any time. In other words, the communication area is set to 25.0 m. After communication between {i,j} vehicles starts, i|j ,A i|j ,b i|j} was adaptively assigned to the i-th vehicle.

[0141] Vehicles connected by edges are com As the distance increases or the road changes (e.g., from an intersection to a straight road), message passing stops and the variables are deallocated to save memory. min =15.0[m] was used.

[0142] <Reference method> To evaluate NOS, a theoretical DNN architecture based on discretized ODE(2), we prepared three reference systems.

[0143] As the first reference system, we used SUMO's native collision-free traffic control with appropriate parameter settings (TraCI speed mode: 31). However, given that SUMO is designed to be controlled using centralized traffic signals, SUMO-controlled vehicles often stopped before entering the intersection to avoid collisions.

[0144] To improve this situation under the collision-free setting of SUMO, we constructed a Plug-and-Play (PnP)Net as a second reference system. PnP-Net uses state transitions implemented with a recursive DNN for updating x. This method avoids analytical derivations as much as possible and is trained in a data-driven manner. The DNN parameter sizes of this PnP-Net were initially set as close as possible to the values ​​of the parameters of NOS.

[0145] As a third reference system, we present trainable DNN functions in both the PRS-Net and DRS-Net varieties.

[0146]

number

[0147] <Implementation> We built the software to run on a server with a CPU (AMD EPYC 7262 8-Core 4) and a GPU (NVIDIA DGX A100 8). We used PyTorch (1.9.0+cu111) for the DNN. The average calculation time for each round of PRS-Net was 45.4 seconds, excluding SUMO-GUI.

[0148] <Experimental Results> The experimental results are shown in Figure 14. Figure 14 shows that the average normalized speed changes with increasing number of rounds. - x av =0.98) was obtained by PRS-Net, and the second fastest ( - x av = 0.97) was obtained with DRS-Net. The performance difference between them is small. That is, both methods are effective for learning associative dynamics based on NOS. In each learning round, we try to maintain the following distance as much as possible. ― We confirmed that θ has been updated.

[0149] Among the reference methods, PnP-Net has the highest speed ( - x av = 0.64), but the NOS score could not be reached because the additional cost of maintaining distance interfered with training to obtain the desired dynamics.

[0150] For the remaining unlearnable reference methods, vehicles often stopped before entering the intersection to avoid collisions, and the average normalized speed was 0.30 for the unlearnable PRS-Net, 0.31 for the unlearnable DRS-Net, and 0.56 for SUMO's native traffic rectification system. Experimental results showed that NOS-based associative dynamics learning is effective.

[0151] Figure 15 more clearly illustrates the difference between the NOS-based associative dynamics learning (proposed method) and the conventional method (SUMO implementation). As shown in Figure 15, the normalization speed of the proposed method approaches the target value as learning progresses, and the speed is significantly improved compared to the conventional method (SUMO implementation).

[0152] In addition, Figure 16 shows the convergence curve for the evaluation set, where the closer it is to 0.0, the better the performance. As shown in Figure 16, the loss value (proportional to the difference between the target value and the current value of the normalized speed) decreases as learning progresses.

[0153] (Device configuration example) FIG. 17 shows an example of the configuration of a vehicle 1. As shown in FIG. 17, the vehicle 1 has a camera 11, a sensor 12, a control device 100, a communication unit 13, and a drive unit 14. The camera 11 acquires an image of the surroundings. The sensor 12 acquires its own position information by, for example, Lidar or GPS. The sensor 12 may include a function to acquire its own speed. Furthermore, a sensor that acquires information other than the above may be mounted.

[0154] The control device 100 receives external information acquired by the camera 11 and the sensor 12, processes the algorithm 1, and outputs a state (velocity) x and a message z. The control device 100 may also include a parameter learning function using back propagation based on equation (11).

[0155] The communication unit 13 receives messages transmitted from other adjacent vehicles and passes the messages to the control device 100, and also transmits messages output from the control device 100 to other adjacent vehicles. Furthermore, when learning is performed by the control server 200, the communication unit 13 transmits the recorded data {x, o, A, b} obtained in the state update to the control server 200, and receives the latest learned parameter θ from the control server 200.

[0156] The drive unit 14 includes a function (engine, motor, etc.) for running the vehicle in accordance with the state x output from the control device 100. For example, when a certain speed is output as a state from the control device 100, the drive unit 14 drives the vehicle to run at that speed.

[0157] Fig. 18 shows an example of the configuration of the control device 100. As shown in Fig. 18, the control device 100 includes an input unit 110, a state update unit 120, an output unit 130, a data storage unit 140, and a learning unit 150. Note that when learning is performed by the control server 200, the control device 100 does not necessarily need to include the learning unit 150.

[0158] The input unit 110 inputs external information o acquired by the camera 11 or the sensor 12, and a message z received from an adjacent vehicle. The state update unit 120 is a DNN that realizes an NOS that updates the states of x and z according to Algorithm 1.

[0159] The output unit 130 outputs the state x and message z obtained by the state update unit 120. The data storage unit 140 records the data {x, o, A, b} obtained in the processing process by the state update unit 120 for each discrete time. The data storage unit 140 also stores the latest learned parameter θ, and the state update unit 120 performs state update processing using the latest learned parameter θ.

[0160] The learning unit 150 learns the parameter θ by backward propagation (11) using the recorded data {x, o, A, b}, and stores the learned parameter θ in the data storage unit 140.

[0161] 19 shows an example of the configuration of the control server 200. As shown in FIG. 19, the control server 200 includes an input unit 210, a learning unit 220, an output unit 230, and a data storage unit 240.

[0162] The input unit 210 receives the recorded data {x, o, A, b} from each vehicle. The data storage unit 240 stores the recorded data {x, o, A, b} received from each vehicle. The learning unit 220 learns the parameter θ by backward propagation (11) using the recorded data {x, o, A, b} stored in the data storage unit 240. The output unit 230 transmits the learned parameter θ to each vehicle.

[0163] (Example of hardware configuration) Both the control device 100 and the control server 200 can be realized by, for example, causing a computer to execute a program.

[0164] That is, the device can be realized by executing a program corresponding to the processing performed by the device using hardware resources such as a CPU, GPU, and memory built into a computer. The program can be recorded on a computer-readable recording medium (such as a portable memory) and stored or distributed. The program can also be provided via a network such as the Internet or email.

[0165] Fig. 20 is a diagram showing an example of the hardware configuration of the computer. The computer in Fig. 20 includes a drive device 1000, an auxiliary storage device 1002, a memory device 1003, a processor 1004, an interface device 1005, a display device 1006, an input device 1007, and an output device 1008, all of which are interconnected by a bus BS. The processor 1004 may be a CPU, a GPU, or a combination of a CPU and a GPU.

[0166] A program for realizing processing on the computer is provided by a recording medium 1001 such as a CD-ROM or a memory card. When the recording medium 1001 storing the program is set in the drive device 1000, the program is installed from the recording medium 1001 to the auxiliary storage device 1002 via the drive device 1000. However, the program does not necessarily have to be installed from the recording medium 1001, but may be downloaded from another computer via a network. The auxiliary storage device 1002 stores the installed program as well as necessary files, data, etc.

[0167] The memory device 1003 reads and stores the program from the auxiliary storage device 1002 when an instruction to start the program is received. The processor 1004 realizes the functions related to the light touch maintenance device 100 in accordance with the program stored in the memory device 1003. The interface device 1005 is used as an interface for connecting to a network, etc. The display device 1006 displays a GUI (Graphical User Interface) etc. according to the program. The input device 1007 is composed of a keyboard, mouse, buttons, touch panel, etc., and is used to input various operation instructions. The output device 1008 outputs the calculation results.

[0168] (Summary, Effects) As explained above, in this embodiment, we propose NOS as an extension of NODE for learning associative dynamics, and assign the two subdynamics of Equation (2) to (i) internal state update and (ii) message passing. Furthermore, we constructed NOS-based DNN architectures such as PRS-Net and DRS-Net through discretization of Equation (2) based on the operator decomposition method.

[0169] Also, NOS is a dynamics parameter that maximizes the average speed to the target value. ― It was applied to the signal-free traffic rectification problem, where the objective is to find θ, and its effectiveness was observed.

[0170] This is expected to reduce travel and transportation times to the maximum extent possible without using traffic signals and preventing traffic accidents (collisions).

[0171] (Addendum) This specification discloses at least the traffic control system, control device, control method, and program described in the following sections. (Section 1) A control device in a traffic control system includes a plurality of mobile bodies each having a control device, and the plurality of mobile bodies autonomously perform traffic control so as to prevent collisions between the mobile bodies, a state update unit that updates the state of the moving object based on state update dynamics having sub-dynamics for updating the state of the moving object and sub-dynamics for message passing between the moving object and other moving objects in the vicinity of the moving object under constraints for preventing collisions between the moving objects; an output unit that outputs the state updated by the state update unit and a message; A control device comprising: (Section 2) The state update unit executes a process of alternately repeating the state update of the moving object and the message passing using a neural network. 2. The control device according to claim 1. (Section 3) The parameters in the neural network are updated by backpropagation calculation based on the adjoint method so that the velocity of the moving object approaches the target velocity under the constraints. 3. The control device according to claim 2. (Section 4) The state update unit calculates a state at a next time point based on a state at a certain time point, an external control input at that time point, and a message received from another moving body at that time point. A control device according to any one of claims 1 to 3. (Section 5) The constraint is that the distance between the moving body and the other moving body, which is calculated based on the external control input, must be equal to or greater than a predetermined distance. 5. The control device according to claim 4. (Section 6) A plurality of moving bodies each equipped with the control device according to any one of paragraphs 1 to 5; a control server that updates parameters of the neural network included in the state update unit by backpropagation calculation based on an adjoint method so that an average speed of a plurality of moving bodies approaches a target speed under the constraints; A traffic control system comprising: (Section 7) A control method executed by a control device in a traffic rectification system that includes a plurality of mobile bodies each equipped with a control device, and in which the plurality of mobile bodies autonomously perform traffic rectification so as to prevent collisions between the mobile bodies, comprising: a state updating step of updating the state of the mobile object based on state update dynamics having sub-dynamics for updating the state of the mobile object and sub-dynamics for message passing between the mobile object and other mobile objects in the vicinity of the mobile object, under constraints for preventing collisions between the mobile objects; an output step of outputting the state updated by the state update step and a message; A control method comprising: (Section 8) A program for causing a computer to function as each part of the control device described in any one of paragraphs 1 to 5.

[0172] Although the present embodiment has been described above, the present invention is not limited to such a specific embodiment, and various modifications and changes are possible within the scope of the gist of the present invention described in the claims. [Explanation of symbols]

[0173] 1 vehicle 11 Camera 12 sensors 13 Communications Department 14 Drive unit 100 control device 110 Input section 120 Status Update Section 130 Output section 140 Data storage unit 150 Learning Department 200 Control Server 210 Input section 220 Learning Department 230 Output section 240 Data storage unit 300 Network 1000 Drive Device 1001 Recording media 1002 Auxiliary storage device 1003 Memory device 1004 processor 1005 Interface device 1006 Display device 1007 Input Device

Claims

1. A control device in a traffic control system includes a plurality of mobile bodies each having a control device, and the plurality of mobile bodies autonomously perform traffic control so as to prevent collisions between the mobile bodies, a state update unit that updates the state of the moving object based on state update dynamics having sub-dynamics for updating the state of the moving object and sub-dynamics for message passing between the moving object and other moving objects in the vicinity of the moving object under constraints for preventing collisions between the moving objects; an output unit that outputs the state updated by the state update unit and a message; the state update unit executes a process of alternately repeating the state update of the moving object and the message passing by using a neural network model for solving an initial value problem of an ordinary differential equation that expresses the state update dynamics; The parameters in the neural network are updated by backpropagation calculation based on the adjoint method so that the velocity of the moving object approaches the target velocity under the constraints. Control device.

2. The state update unit calculates a state at a next time point based on a state at a certain time point, an external control input at that time point, and a message received from another moving body at that time point. The control device according to claim 1 .

3. The constraint is that the distance between the moving body and the other moving body, which is calculated based on the external control input, must be equal to or greater than a predetermined distance. The control device according to claim 2 .

4. A plurality of moving bodies each equipped with the control device according to any one of claims 1 to 3; a control server that updates parameters of the neural network included in the state update unit by backpropagation calculation based on an adjoint method so that an average speed of a plurality of moving bodies approaches a target speed under the constraints; A traffic control system comprising:

5. A control method executed by a control device in a traffic rectification system that includes a plurality of mobile bodies each equipped with a control device, and in which the plurality of mobile bodies autonomously perform traffic rectification so as to prevent collisions between the mobile bodies, comprising: a state updating step of updating the state of the mobile object based on state update dynamics having sub-dynamics for updating the state of the mobile object and sub-dynamics for message passing between the mobile object and other mobile objects in the vicinity of the mobile object, under constraints for preventing collisions between the mobile objects; an output step of outputting the state updated by the state update step and a message; In the state updating step, the control device executes a process of alternately repeating the state update of the moving object and the message passing by using a neural network model for solving an initial value problem of an ordinary differential equation that represents the state update dynamics; The parameters in the neural network are updated by backpropagation calculation based on the adjoint method so that the velocity of the moving object approaches the target velocity under the constraints. Control method.

6. A program for causing a computer to function as each unit in the control device according to any one of claims 1 to 3.

Images