A multi-agent reinforcement learning method for vehicle network computation offloading

Abstract

The present invention discloses a multi-agent reinforcement learning method for offloading computation in an Internet of Vehicles (IoV), comprising the following steps: step S1, constructing a system model; step S2, constructing a communication model; step S3, constructing a task and computation model; step S4, constructing an energy consumption model; step S5, constructing an optimization model and establishing a Markov decision process; step S6, using the MADDPG algorithm to perform neural network training on the optimization model; and step S7, deploying the trained network to each agent. The present invention uses the service vehicles and MEC servers in the IoV system as agents in reinforcement learning to make resource allocation decisions, and uses multi-agent reinforcement learning to allow each agent to obtain its own strategy network. After training, each agent can quickly output its own actions based on current local information without excessive communication, thus being able to cope with the rapidly changing IoV environment and reduce the system's energy consumption.

Description

Technical Field

[0001] The present application belongs to the field of vehicle network communication technology, and specifically relates to a multi-agent reinforcement learning vehicle network computing offloading method. Background Art

[0002] As the automotive industry becomes increasingly intelligent, intelligent connected vehicles (ICVs) have become a key pillar of global automotive industry development. With the advancement of network technology and the emergence of advanced technologies such as autonomous and driverless driving, existing automotive application systems are no longer able to meet the growing demand for in-vehicle applications. Just as mobile phones offer a variety of services and applications beyond text messaging and calls, cars have evolved into intelligent mobile spaces, powered by cloud computing and offering a wide range of functional, safety, and entertainment applications. However, the distance between the cloud and the vehicle is significant. To access cloud services, users must perform multiple network hops, resulting in slow responses. Furthermore, data from the vehicle must be uploaded through the core network, which is not only costly but also causes network congestion. Ultimately, this leads to degraded service quality due to severe latency and unstable transmission.

[0003] Mobile Edge Computing (MEC) technology can provide computing resources, storage resources, and substantial network bandwidth close to user terminals, shortening data transmission distances while also protecting data security and user privacy. One of the key applications of MEC in the Internet of Vehicles (IoV) is computation offloading. When a vehicle faces computationally intensive tasks that are difficult for the onboard controller to handle, it can offload these tasks to the MEC for processing. Current research primarily considers offloading tasks to the MEC or other vehicles, while research on simultaneous offloading to both vehicles and the MEC is limited. Hybrid offloading, on the other hand, can directly offload some tasks to idle vehicles for computation via a vehicle-to-vehicle approach, alleviating communication pressure between the MEC and a large number of vehicles. However, traditional optimization methods often require obtaining a large amount of global information encompassing all vehicles and the MEC. In an IoV environment, this global information is constantly changing due to rapidly changing scenarios, and the resulting communication overhead can affect the speed of offloading decisions. Summary of the Invention

[0004] The present invention discloses a multi-agent reinforcement learning vehicle network computing offloading method, which can effectively solve the technical problems involved in the background technology.

[0005] To achieve the above object, the technical solution of the present invention is:

[0006] A multi-agent reinforcement learning vehicle network computing offloading method includes the following steps:

[0007] Step S1: Build a system model, including MEC and intelligent connected vehicles, and regard them as intelligent agents. Then, intelligent connected vehicles are divided into customer vehicles and service vehicles based on resource usage. Customer vehicles offload tasks to MEC and service vehicles.

[0008] Step S2: Build a communication model to obtain the transmission efficiency of customer vehicles, service vehicles, and MEC;

[0009] Step S3: Build a task and calculation model to obtain the total task offloading delay of the system;

[0010] Step S4: construct an energy consumption model to obtain the total energy consumption of the system;

[0011] Step S5: construct an optimization model based on the total task offloading delay and total energy consumption and establish a Markov decision process; wherein the optimization model is:

[0012]

[0013]

[0014]

[0015]

[0016]

[0017]

[0018] In the formula, H represents the optimization target, C represents the customer vehicle set, β t represents the delay ratio coefficient, represents the total task offloading delay, β E represents the energy consumption proportional coefficient, represents the total energy consumption, a c,m,s represents the offloading action of the client vehicle offloading the task to the service vehicle and MEC, a c represents the total unloading action of the task vehicle, b 0,s,c represents the computing resources allocated by the service vehicle to the customer vehicle, B S represents the maximum computing resources of the service vehicle, b m,0,c represents the computing resources allocated by the server to the client vehicle, B M Indicates the maximum computing resources of MEC, p c，m，s represents the task offloading power;

[0019] The Markov decision process specifically includes:

[0020] Step S51: Define the state space of the customer vehicle and action space The expressions are:

[0021]

[0022]

[0023] Where, represents the channel gain, b t-1 Indicates the computing resource allocation action of the devices that the customer vehicle can connect to, Indicates the task data size, Indicates the current speed of the customer's vehicle, P c Indicates unloading power;

[0024] Step S52: Define the state space of the service vehicle and action space The expressions are:

[0025]

[0026]

[0027] Where, Indicates the current speed of the service vehicle, Indicates the remaining resources of the service vehicle at the previous moment;

[0028] Step S53: Define the state space of MEC and action space The expressions are:

[0029]

[0030]

[0031] Where, Indicates the remaining resources of MEC at the previous moment;

[0032] Step S54: Define reward function r t , whose expression is:

[0033]

[0034] Step S6: Use the MADDPG algorithm to perform neural network training on the optimization model;

[0035] Step S7: Deploy the trained network to each agent.

[0036] As a preferred improvement of the present invention, in step S1, the MEC is represented by the set M = {1, 2, ..., m}, the customer vehicles are represented by the set C = {1, 2, ..., c}, and the service vehicles are represented by the set S = {1, 2, ..., s}.

[0037] As a preferred improvement of the present invention, in step S2, orthogonal frequency division multiple access technology is used for information transmission. The customer vehicle first obtains the channel gain by connecting the MEC and the server to calculate the spectrum efficiency, and then obtains the transmission efficiency according to the Shannon formula.

[0038] As a preferred improvement of the present invention, in step S3, the task and calculation model include:

[0039]

[0040]

[0041]

[0042]

[0043]

[0044] Where, Indicates the task upload delay, a c，m，s represents the unloading ratio of customer vehicles, Indicates the amount of data transferred by the offload task. Indicates the transmission rate, represents edge computing latency, b m，s，c represents the computing resources allocated by MEC and service vehicles to customer vehicles, Indicates the maximum delay in all edge computing. Indicates the local computing delay, a c，0，0 represents the proportion of customer vehicles remaining in the local area, b 0,0,c represents the computing resources allocated to the local part of the task by the client vehicle, Indicates the total task offloading latency.

[0045] As a preferred improvement of the present invention, in step S4, the energy consumption model includes:

[0046]

[0047]

[0048]

[0049]

[0050] Where, represents the transmission energy consumption of the customer's vehicle, p c，m，s represents the unloading power, represents the local computing energy consumption of the customer vehicle, ξ represents the energy consumption coefficient of the vehicle computing chip, represents the calculated energy consumption of the service vehicle, a c,0,s represents the ratio of tasks unloaded from customer vehicles to service vehicles, b 0,s,c represents the computing resources allocated by the service vehicle to the customer vehicle, Indicates the total energy consumption.

[0051] As a preferred improvement of the present invention, step S6 specifically includes:

[0052] Step S61: The customer vehicle, service vehicle, and MEC are all considered as intelligent agents, each of which includes an Actor network, a Critic network, and its corresponding target network;

[0053] Step S62: The client vehicle outputs an unloading action based on its current state and broadcasts it to the MEC and the service vehicle;

[0054] Step S63: MEC and service vehicle receive the unloading action and calculate the unloading according to their own assigned actions, and all agents enter the next state;

[0055] Step S64: storing the current state, current action, reward, and next state of each agent in the experience replay pool as samples for central control to extract and train;

[0056] Step S65: The critic network uses the state and action of the agent in the sample to obtain the corresponding action value;

[0057] Step S66: Each agent updates the Actor network parameters according to the action value returned by the central controller;

[0058] Step S67: Update Critic network parameters;

[0059] Step S68: Update the target network parameters until convergence.

[0060] As a preferred modification of the present invention, in step S7, the trained Actor network is deployed on the intelligent agent. The customer vehicle makes an unloading decision based on the current local information, and then broadcasts it to the service vehicle and MEC. The service vehicle and MEC then combine the unloading decision and their own observed information to make an allocation strategy.

[0061] The beneficial effects of the present invention are as follows:

[0062] 1. The method proposed in this paper is suitable for vehicle networking scenarios where vehicles have a large number of computationally intensive tasks. By using MADDPG to solve the optimization problem proposed in this paper, the problem of insufficient vehicle computing power can be effectively alleviated.

[0063] 2. The method proposed in this invention takes into account the idle resources of vehicles in MEC and the Internet of Vehicles system. It not only offloads tasks to MEC servers, but also improves the efficiency of task offloading to customer vehicles and fully utilizes system resources to achieve load balancing.

[0064] 3. By treating service vehicles and MEC servers as agents in reinforcement learning to make resource allocation decisions, and using multi-agent reinforcement learning to allow each agent to obtain its own strategy network, after training, each agent can quickly output its own actions based on the current local information without excessive communication, ensuring that the task is completed within the delay constraint. In addition, the computing resource allocation of MEC and service vehicles is optimized, which can cope with the rapidly changing Internet of Vehicles environment and reduce the system's energy consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0065] Figure 1 This is a flowchart of a multi-agent reinforcement learning vehicle network computing offloading method of the present invention. DETAILED DESCRIPTION

[0066] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the embodiments of the present invention. Obviously, the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0067] It should be noted that all directional indications in the embodiments of the present invention (such as up, down, left, right, front, back, etc.) are only used to explain the relative position relationship, movement status, etc. between the various components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.

[0068] In addition, the terms "first," "second," and so on, used in this disclosure are for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referenced. Thus, a feature specified as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of this disclosure, "plurality" means at least two, such as two or three, unless otherwise specifically defined.

[0069] In the present invention, unless otherwise specified or limited, the terms "connection" and "fixation" should be understood in a broad sense. For example, "fixation" can mean fixed connection, detachable connection, or integration; mechanical connection or electrical connection; direct connection or indirect connection through an intermediate medium; internal communication between two elements or interaction between two elements, unless otherwise specified. Those skilled in the art will be able to understand the specific meanings of the above terms in the present invention based on specific circumstances.

[0070] In addition, the technical solutions between the various embodiments of the present invention can be combined with each other, but it must be based on the fact that ordinary technicians in this field can implement it. When the combination of technical solutions is mutually contradictory or cannot be implemented, it should be deemed that such a combination of technical solutions does not exist and is not within the scope of protection required by the present invention.

[0071] See also Figure 1 As shown, the present invention proposes a multi-agent reinforcement learning vehicle network computing offloading method, which includes the following steps:

[0072] Step S1: Construct a system model.

[0073] The network contains multiple MECs and intelligent connected vehicles. Assume that MECs are evenly distributed along the roadside within a certain range and are connected to each other via high-speed fiber optic links. All intelligent connected vehicles are within the MEC range. Intelligent connected vehicles are categorized into client vehicles and service vehicles based on resource usage. MECs are represented by the set M = {1, 2, ..., m}, client vehicles by the set C = {1, 2, ..., c}, and service vehicles by the set S = {1, 2, ..., s}. Client vehicles can connect to MECs, to nearby service vehicles, or to more distant service vehicles through MECs. However, due to network latency and packet loss associated with multiple hops, offloading tasks to more distant service vehicles is not considered.

[0074] Step S2: Construct a communication model.

[0075] In order to avoid mutual interference between channels used by multiple vehicles, this application uses orthogonal frequency division multiple access technology to transmit information. Before the algorithm is executed, the client vehicle can obtain the channel gain by connecting the service vehicle and the server, where the channel gain between the client vehicle and the service vehicle at time t is expressed as h c,s,0 , the channel gain between the client vehicle and MEC at time t is expressed as h c,0,m , we can get the spectrum efficiency at time t as expressed in formula (1) Then, according to Shannon's formula, the transmission rate at time t is obtained As shown in formula (2):

[0076]

[0077]

[0078] Where p c,m.s represents the transmission power between the client vehicle, the service vehicle and the server, σ 2 is the noise power, B c Indicates the bandwidth between the service vehicle and MEC.

[0079] Step S3: Construct task and calculation model.

[0080] There is a task management module on the customer's vehicle, which generates tasks in each time slot Including task data size And task delay constraints Therefore, the task at the current time t can be represented by a tuple express.

[0081] The task can be split and processed. To meet the delay requirement, the task can be partially offloaded to the MEC or nearby service vehicles for calculation. The offloading action a of each vehicle can be obtained. c =(a c,m,s ,a c,0,0 )=(a c,1,0 ,...,a c,m,0 ,a c,0,1 ,...,a c,0,S ,a c,0,0 ), where a c,m,0 represents the proportion of tasks unloaded from customer vehicle c to the mth MEC, a c,0,s represents the proportion of tasks unloaded from customer vehicle c to the s-th service vehicle, a c，0，0 represents the proportion of vehicles remaining in local calculations, and the power selection action can be expressed as p c =(p c,1,0 ,,...,p c,M,0 ,p c,0,1 ,...,p c,0,S ), where p c,m,0 and p c,0,s They represent the power selected when offloading the task to the mth server and the sth service vehicle, respectively. From this, the task upload delay can be obtained, as shown in formula (3):

[0082]

[0083] The allocation action for the service vehicle providing unloading service is represented by b s =(b 0,1,c ,...,b 0,S,c), the allocation action of MEC is represented as b m =(b 1,0,c ,...,b M,0,c ), so b=(b m,s,c ,b 0,0,c )=(b m ,b s ,b 0,0,c ), where b 0,0,c represents the computing resources allocated to the local part of the task in the customer vehicle, from which the time delay of the task for edge computing can be obtained, as shown in formula (4):

[0084]

[0085] Since tasks can be computed in parallel, the total delay of the edge computing task is the maximum delay of all edge computing tasks, as shown in formula (5):

[0086]

[0087] The calculation delay of local execution of some tasks is expressed as formula (6):

[0088]

[0089] In summary, the total task offloading delay can be calculated as shown in formula (7):

[0090]

[0091] Step S4: Construct an energy consumption model.

[0092] The energy consumption required for edge computing includes the energy consumption of customer vehicle transmission tasks and the energy consumption of service vehicles and MEC computing. Because MEC has a stable power supply, its computing energy consumption can be ignored and only the energy consumption of the vehicle can be considered.

[0093] Among them, the energy consumption of the customer vehicle transmission task is shown in formula (8), the local computing energy consumption of the customer vehicle is shown in formula (9), and the computing energy consumption of the service vehicle is shown in formula (10):

[0094]

[0095]

[0096]

[0097] Where ξ represents the energy consumption coefficient of the vehicle computing chip, b 0,s,c Represents the computing resources allocated by the service vehicle to the customer vehicle.

[0098] Thus, the total energy consumption of the system can be obtained as shown in formula (11):

[0099]

[0100] Step S5: construct an optimization model based on the total task offloading delay and total energy consumption and establish a Markov decision process, wherein the optimization model is:

[0101]

[0102]

[0103]

[0104]

[0105]

[0106] In the formula, H represents the optimization target, β t Denotes the delay ratio coefficient, β E Indicates the energy consumption proportional coefficient, B S represents the maximum computing resources of the service vehicle, B M Represents the maximum computing resources of MEC, P max Indicates maximum power.

[0107] The Markov decision process specifically includes:

[0108] Step S51: First define the state space of the customer vehicle and action space Its expressions are shown in formula (13) and formula (14) respectively:

[0109]

[0110]

[0111] Where, represents the channel gain, b t-1 Indicates the computing resource allocation action of the devices that the customer vehicle can connect to, Indicates the task, Indicates the current speed of the customer's vehicle;

[0112] Step S52: Define the state space of the service vehicle and action space Its expressions are shown in formula (15) and formula (16) respectively:

[0113]

[0114]

[0115] Where, Indicates the current speed of the service vehicle, Indicates the remaining resources of the service vehicle at the previous moment.

[0116] Step S53: Define the state space of MEC and action space Its expressions are shown in formula (17) and formula (18) respectively:

[0117]

[0118]

[0119] Where, Indicates the remaining resources of MEC at the previous moment;

[0120] Step S54: Define reward function r t , as shown in formula (19):

[0121]

[0122] The reward function can reduce the delay and energy consumption and meet the optimization conditions.

[0123] Step S6: Use the MADDPG algorithm to perform neural network training on the optimization model, specifically including:

[0124] Step S61: The customer vehicle, service vehicle, and MEC are all considered as intelligent agents, and each intelligent agent deploys its own Actor network π locally. c,m,s , initialize the network parameters to θ c ,θ s ,θ m ; Deploy the Critic network corresponding to each agent in the central controller The initialization parameter is ω c ,ω m ,ω s , and initialize the target network of each neural network To alleviate the overestimation problem caused by bootstrapping;

[0125] Step S62: The customer's vehicle is based on its current state Output unloading action And broadcast it to MEC and service vehicles;

[0126] Step S63: The MEC and the service vehicle determine their respective current states after receiving the unloading action. as well as and broadcast their respective allocation actions After obtaining the unloading strategy and allocation strategy, the calculation unloading is performed and all agents enter the next state

[0127] Step S64: The current state of each agent Current Action award Next state And the next step Send it to the central controller and The data is stored in the experience replay pool. When the number of samples in the experience replay pool reaches a certain amount, the central controller starts to extract N samples from it for training.

[0128] Step S65: The critic network uses the state and action of the agent in the sample to obtain the corresponding action value, as shown in formula (20):

[0129]

[0130] Step S66: The central controller returns the action value to each agent, and each agent calculates its own gradient to update its own Actor network parameters:

[0131]

[0132] Step S67: The central controller minimizes the loss function Δ t Update the critic network parameters of each agent to avoid overestimating the target value calculated using the target network parameters:

[0133]

[0134] in, It represents the predicted value at the next moment, which is expressed by formula (23):

[0135]

[0136] Step S68: Periodically update target network parameters as well as And continue training until convergence.

[0137] Step S7: Deploy the trained actor network to each agent. At this point, a central controller and critic network are no longer required; agents can make decisions based on their own local observations. First, the client vehicle makes an offloading decision based on current local information. This decision is then broadcast to the service vehicle and MEC server. The service vehicle and MEC server then combine the offloading decision with their own observations to formulate a distribution strategy.

[0138] The beneficial effects of the present invention are as follows:

[0139] 1. The method proposed in this paper is suitable for vehicle networking scenarios where vehicles have a large number of computationally intensive tasks. By using MADDPG to solve the optimization problem proposed in this paper, the problem of insufficient vehicle computing power can be effectively alleviated.

[0140] 2. The method proposed in this invention takes into account the idle resources of vehicles in MEC and the Internet of Vehicles system. It not only offloads tasks to MEC servers, but also improves the efficiency of task offloading to customer vehicles and fully utilizes system resources to achieve load balancing.

[0141] 3. By treating service vehicles and MEC servers as agents in reinforcement learning to make resource allocation decisions, and using multi-agent reinforcement learning to allow each agent to obtain its own strategy network, after training, each agent can quickly output its own actions based on the current local information without excessive communication, ensuring that the task is completed within the delay constraint. In addition, the computing resource allocation of MEC and service vehicles is optimized, which can cope with the rapidly changing Internet of Vehicles environment and reduce the system's energy consumption.

[0142] The embodiments of the present application are described above in conjunction with the accompanying drawings, but the present application is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of this application, ordinary technicians in this field can also make many forms without departing from the purpose of this application and the scope of protection of the claims, all of which are within the protection of this application.

Claims

1. A multi-agent reinforcement learning vehicle network computing offloading method, characterized by: The following steps are involved: Step S1: Build a system model, including MEC and intelligent connected vehicles, and regard both MEC and intelligent connected vehicles as intelligent entities. Then, intelligent connected vehicles are divided into customer vehicles and service vehicles based on resource usage. Customer vehicles offload tasks to MEC and service vehicles. Step S2: Build a communication model to obtain the transmission efficiency of customer vehicles, service vehicles, and MEC; Step S3: Build a task and calculation model to obtain the total task offloading delay of the system; Step S4: construct an energy consumption model to obtain the total energy consumption of the system; Step S5: construct an optimization model based on the total task offloading delay and total energy consumption and establish a Markov decision process; wherein the optimization model is: In the formula, H represents the optimization target, C represents the customer vehicle set, β t represents the delay ratio coefficient, represents the total task offloading delay, β E represents the energy consumption proportional coefficient, represents the total energy consumption, a c,m,s represents the offloading action of the client vehicle offloading the task to the service vehicle and MEC, a c represents the total unloading action of the task vehicle, b 0,s,c represents the computing resources allocated by the service vehicle to the customer vehicle, B S represents the maximum computing resources of the service vehicle, b m,0,c represents the computing resources allocated by the server to the client vehicle, B M Indicates the maximum computing resources of MEC, p c，m，s represents the task offloading power; The Markov decision process specifically includes: Step S51: Define the state space of the customer vehicle and action space The expressions are: Where, represents the channel gain, b t-1 Indicates the computing resource allocation action of the devices that the customer vehicle can connect to, Indicates the task data size, Indicates the current speed of the client's vehicle; Step S52: Define the state space of the service vehicle and action space The expressions are: Where, Indicates the current speed of the service vehicle, Indicates the remaining resources of the service vehicle at the previous moment; Step S53: Define the state space of MEC and action space The expressions are: Where, Indicates the remaining resources of MEC at the previous moment; Step S54: Define reward function r t , whose expression is: Step S6: Use the MADDPG algorithm to perform neural network training on the optimization model; Step S7: Deploy the trained network to each agent.

2. The method according to claim 1, characterized in that In step S1, the MEC is represented by a set M = {1, 2, ..., m}, the customer vehicles are represented by a set C = {1, 2, ..., c}, and the service vehicles are represented by a set S = {1, 2, ..., s}.

3. The method according to claim 1, characterized in that In step S2, orthogonal frequency division multiple access technology is used for information transmission. The customer vehicle first obtains the channel gain by connecting to the MEC and the server to calculate the spectrum efficiency, and then obtains the transmission efficiency according to the Shannon formula.

4. The method according to claim 1, wherein In step S3, the task and calculation model include: Where, Indicates the task upload delay, Indicates the amount of data transferred by the offload task. Indicates the transmission rate, represents edge computing latency, b m，s，c represents the computing resources allocated by MEC and service vehicles to customer vehicles, Indicates the maximum delay in all edge computing. Indicates the local computing delay, a c，0，0 represents the proportion of customer vehicles remaining in the local area, b 0,0,c represents the computing resources allocated to the local part of the task by the client vehicle, Indicates the total task offloading latency.

5. The method according to claim 1, wherein In step S4, the energy consumption model includes: Where, represents the transmission energy consumption of the customer's vehicle, p c，m，s represents the unloading power, represents the local computing energy consumption of the customer vehicle, ξ represents the energy consumption coefficient of the vehicle computing chip, represents the calculated energy consumption of the service vehicle, a c,0,s represents the ratio of tasks unloaded from customer vehicles to service vehicles, b 0,s,c represents the computing resources allocated by the service vehicle to the customer vehicle, Indicates the total energy consumption.

6. The method according to claim 1, wherein Step S6 specifically includes: Step S61: The customer vehicle, service vehicle, and MEC are all considered as intelligent agents, each of which includes an Actor network, a Critic network, and its corresponding target network; Step S62: The client vehicle outputs an unloading action based on its current state and broadcasts it to the MEC and the service vehicle; Step S63: MEC and service vehicle receive the unloading action and calculate the unloading according to their own assigned actions, and all agents enter the next state; Step S64: storing the current state, current action, reward, and next state of each agent in the experience replay pool as samples for central control to extract and train; Step S65: The critic network uses the state and action of the agent in the sample to obtain the corresponding action value; Step S66: Each agent updates the Actor network parameters according to the action value returned by the central controller; Step S67: Update Critic network parameters; Step S68: Update the target network parameters until convergence.

7. The method according to claim 1, characterized in that In step S7, the trained Actor network is deployed on the agent. The client vehicle makes an unloading decision based on the current local information, and then broadcasts it to the service vehicle and MEC. The service vehicle and MEC then combine the unloading decision and their own observed information to make an allocation strategy.

Images