Complaint processing method, system and device based on reinforcement learning and storage medium
By employing a reinforcement learning-based complaint handling method, utilizing Markov decision processes and deep Q-network algorithms, complaint solutions are automatically generated, addressing the issues of low dynamic adaptability and resource efficiency in existing technologies, and achieving efficient and flexible complaint decision-making.
Patent Information
- Application Number
- CN202511153583.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-08-18
- Publication Date
- 2025-12-02
AI Technical Summary
Existing complaint handling solutions lack dynamic adaptability and feedback learning capabilities, are inefficient in terms of computing resources and data utilization, and are difficult to generate optimal decisions quickly in complex environments.
A reinforcement learning-based approach is used to model resident complaint data as a Markov decision process. A complaint decision model is constructed using a deep Q-network and a proximal policy optimization algorithm. Resident satisfaction and resource utilization are used as reward functions to automatically generate solutions.
It achieves adaptability and flexibility in complaint handling, improves response efficiency, reduces reliance on human resources, and enhances the accuracy of decision-making and the level of system automation.
Smart Images
Figure CN121052297A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to reinforcement learning technology, and more specifically to a method, system, device, and storage medium for handling complaints based on reinforcement learning. Background Technology
[0002] RL (Reinforcement Learning): A machine learning method that uses an agent to interact with the environment and receive feedback to optimize policies with the goal of maximizing cumulative rewards, thereby autonomously learning the optimal decision path.
[0003] DQN (Deep Q Network): A deep reinforcement learning algorithm that applies deep neural networks to Q-learning, guiding decision-making by estimating the expected reward value for each state-action pair.
[0004] PPO (Proximal Policy Optimization): A policy optimization algorithm that maintains policy continuity by limiting update magnitude, effectively improving the stability and training efficiency of reinforcement learning models.
[0005] MDP (Markov Decision Process): A mathematical model used to describe decision problems, in which each state transition depends only on the current state and action, and is independent of past decisions, thus conforming to the Markov property.
[0006] State Space: The set of all possible states used by the model in the decision-making process, which usually includes indicators from multiple dimensions such as resource allocation, resident satisfaction, and historical data.
[0007] Action Space: The set of all governance measures that an agent can take in a given state, such as adjusting manpower allocation, optimizing resources, and responding to emergencies.
[0008] Reward Mechanism: A system that provides positive and negative feedback to reinforcement learning models, giving corresponding rewards or penalties based on the decision-making effect (such as resident satisfaction, resource utilization, etc.) to guide the model to optimize its strategy.
[0009] Q-Learning: A model-free reinforcement learning algorithm that selects the optimal decision by estimating the Q-value of each state-action pair. It is often used in discrete action spaces.
[0010] OpenAI Gym: An open-source platform for developing and comparing reinforcement learning algorithms, providing various environments to support the training and testing of reinforcement learning models.
[0011] Existing complaint handling solutions are mostly based on traditional experience-based decision-making and static data analysis, relying primarily on historical data, statistical methods, and expert experience.
[0012] Problems with existing technologies: lack of dynamic adaptability: traditional complaint handling decisions often rely on static rules or expert experience, which are difficult to adapt to real-time changing environmental needs and limit processing efficiency.
[0013] Insufficient ability to learn from feedback in complex environments: Existing technologies are mainly based on fixed statistical models or expert rules, lacking the ability to learn and optimize from actual feedback. This inability to continuously gain experience and optimize in real-world scenarios leads to poor performance of complaint handling measures in new problems or changing environments.
[0014] Inefficient use of computing resources: In existing solutions, rule-guided and manual decision analysis rely on a large amount of human and hardware resources, but lack automated optimization techniques, resulting in low computing efficiency, excessive manual intervention, and difficulty in quickly achieving optimal configuration in emergency events.
[0015] Low data utilization efficiency: Existing solutions mostly rely on simple regression analysis of historical data, lacking response and feedback optimization of real-time data, resulting in insufficient adaptability and accuracy of decision results under multi-dimensional conditions. Summary of the Invention
[0016] To address the shortcomings of existing technologies, this invention provides a method, system, device, and storage medium for handling complaints based on reinforcement learning.
[0017] To achieve the above objectives, the present invention adopts the following technical solution:
[0018] A complaint handling method based on reinforcement learning includes the following steps:
[0019] Historical resident complaint data and corresponding solutions are collected and modeled as a Markov decision process. Based on the historical resident complaint data and corresponding solutions, a state space and an action space are defined. The state space includes the complaint type, priority, and processing progress, while the action space includes actions such as increasing manpower, extending working hours, and increasing response priority.
[0020] Reinforcement learning is performed in a set Markov decision environment using deep learning algorithms. After receiving complaint data in the state space each time, the best response is selected in the action space. A reward function is set according to the governance objectives, based on resident satisfaction and resource utilization indicators. The reward function is adjusted according to resident feedback and resource utilization. Training stops when the model reaches the expected accuracy, and the complaint decision model is obtained.
[0021] When the complaint decision model receives a new complaint, it automatically generates a solution.
[0022] To optimize the above technical solution, the specific measures also include:
[0023] Furthermore, the state space includes complaint type, priority, and processing progress, and the action space includes actions such as increasing manpower, extending working hours, and increasing response priority.
[0024] The specific process of modeling historical resident complaint data and corresponding solutions as a Markov decision process is as follows:
[0025] The state space is structured as follows: complaint categories are mapped to discrete state vectors using one-hot encoding, priorities are quantized into integer state values with a value range of 1-5 corresponding to [urgent, high, medium, low, normal], and processing progress is converted into continuous state values.
[0026] The action space is discretized as follows: adding manpower is subdivided into: {no addition, add 1 person, add 2 people}; extending working time is quantified into: {no extension, extend by 0.5 hours, extend by 1 hour}; and response priority is adjusted into: {maintain, upgrade by 1 level, upgrade by 2 levels}.
[0027] Construct a state transition probability matrix based on historical data, specifically by calculating the probability of transitioning from the current state to another state, as shown in the following formula:
[0028]
[0029] In the formula, P(s) ′ |s,a) indicates that the action a is performed to transition from the current state s to another state s. ′ The probability, N(s,a,s) ′ () indicates that action a is performed from the current state s to another state s. ′ The historical count of the number of times action a is executed in the current state s; N(s,a) represents the historical count of the number of times action a is executed in the current state s.
[0030] Combine the state transition probabilities of each state into a state transition probability matrix;
[0031] Design a reward function.
[0032] Furthermore, the deep learning algorithm is a deep Q-network algorithm, and the specific process of reinforcement learning in a given Markov decision environment using the deep Q-network algorithm is as follows:
[0033] Construct an online Q-network and a target Q-network. The online Q-network uses a three-layer fully connected neural network; the target Q-network has the same structure as the online Q-network.
[0034] Initialize the experience replay pool D with samples (s, a, r, s'), where s represents the current state, a represents the action, r represents the reward, and s' represents the next state;
[0035] Obtain the current state from the state space, select the action corresponding to the state from the action space based on the ε-greedy policy; calculate the reward value of the current action and update the state; put the four-tuple of state, action, reward value and updated state into the experience replay pool;
[0036] Sample a batch of experience from the experience replay pool to update the Q value of the target Q network;
[0037] The loss function of the online Q-network is calculated using the updated Q-value of the target Q-network, and the parameters of the online Q-network are updated based on the gradient of the loss function; each In the next iteration, the parameters of the target Q network are updated by copying the parameters of the online Q network to the target Q network;
[0038] Once the model reaches the expected accuracy, training stops, and the complaint decision model is obtained; otherwise, training continues.
[0039] Furthermore, the deep learning algorithm is a proximal policy optimization algorithm. The specific process of performing reinforcement learning in a given Markov decision environment using the proximal policy optimization algorithm is as follows:
[0040] Construct an Actor-Critic network architecture, comprising an Actor network and a Critic network. States are collected from the state space and input into both the Actor and Critic networks respectively. The Actor network outputs the probability distribution of actions, and the Critic network outputs the estimated value of the states. Policy optimization employs a truncated objective function.
[0041] Calculate the reward value of the action, calculate the TD residual based on the estimated value and the reward value of the action, calculate the GAE advantage based on the TD residual, and construct the truncated objective function based on the GAE advantage and the probability distribution of the action.
[0042] The parameters of the Actor network and Critic network are continuously updated in the direction of gradient descent of the truncated objective function until the network converges, thus obtaining the complaint decision model.
[0043] Furthermore, the specific steps for setting the reward function based on the governance objective are as follows:
[0044] The reward function is designed as follows:
[0045] r = α·Sat + β·Util - γ·Cost
[0046] Where r represents the reward value, Sat represents the satisfaction reward, α is the weight of the satisfaction reward, Sat=2*exp(-0.5*t_delay)+0.3*rating, t_delay represents the portion of the actual processing time that exceeds the standard time, and rating represents the resident evaluation score;
[0047] Util represents the resource utilization reward, β is the weight of the resource utilization reward, Util = min(1, processed_tasks ÷ (staff_hours * 0.8)), processed_tasks represents the number of work orders processed per unit time, and staff_hours represents the total manpower hours invested;
[0048] Cost represents the penalty for resource consumption costs, and γ is the weight of the penalty for resource consumption costs. Cost = 0.2 * Δstaff + 0.1 * Δhours, where Δstaff represents the increased number of personnel and Δhours represents the extended working hours.
[0049] This invention also proposes a complaint processing system based on reinforcement learning, comprising:
[0050] The data collection module is used to collect historical resident complaint data and corresponding solutions;
[0051] The model training module is used to train a complaint decision-making model based on collected historical resident complaint data and corresponding solutions;
[0052] The decision server is used to deploy the trained complaint decision model, run reinforcement learning algorithms, and generate decision recommendations by processing the collected data in real time.
[0053] The execution device is used to perform actual operations based on the decision recommendations of the decision server.
[0054] The present invention also proposes an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, it implements the reinforcement learning-based complaint processing method described above.
[0055] The present invention also proposes a computer-readable storage medium storing a computer program that causes a computer to perform the reinforcement learning-based complaint processing method described above.
[0056] The beneficial effects of this invention are:
[0057] This invention models resident complaints and corresponding solutions as Markov Decision Processes (MDPs) using reinforcement learning algorithms (such as DQN and PPO), dynamically adjusting decision-making strategies based on real-time data. Compared to traditional rule-based decision-making methods, this approach uses an automated learning model for intelligent optimization, automatically generating the optimal decision solution when a resident complaint occurs. This invention possesses adaptability, enabling real-time responses to complex scenarios and significantly improving response efficiency. It enhances the processing efficiency of public services, reduces reliance on human labor, and improves the accuracy and flexibility of the system's automated decision-making.
[0058] This invention defines specific action spaces for different scenarios, including operational options such as personnel allocation, resource reallocation, and priority adjustment, to adapt to different needs. Unlike existing fixed operational methods, this invention flexibly sets the action space according to different scenarios, ensuring that the model selects the optimal action strategy in various scenarios.
[0059] Customizable action space settings enhance the model's flexibility and applicability, enabling the system to adopt optimal solutions in different public service scenarios, thereby meeting needs more efficiently.
[0060] At the end of each cycle, the system collects and analyzes the execution results, and further adjusts the model based on the exploration and feedback mechanism of reinforcement learning to optimize subsequent decisions. This avoids the limitations of relying on manual optimization in traditional solutions, and continuously adjusts decision-making strategies through an automated feedback mechanism, resulting in continuous improvement in the effectiveness of this invention in real-world scenarios. It achieves self-learning and continuous optimization of the model, enabling the system to continuously adapt and optimize in practice, gradually improving the level of automation and response quality of public services. Attached Figure Description
[0061] Figure 1 Modeling process of Markov decision process in complaint handling. Detailed Implementation
[0062] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.
[0063] Example 1
[0064] This invention proposes a complaint handling method based on reinforcement learning, comprising the following steps:
[0065] Collect historical resident complaint data and corresponding solutions, and create an environment adapted to the data using OpenAI Gym or a custom simulation platform. Model the historical resident complaint data and corresponding solutions as a Markov Decision Process (MDP), and define the state space and action space based on the historical resident complaint data and corresponding solutions. The state space includes the complaint category, priority, and processing progress, while the action space includes actions such as increasing manpower, extending working hours, and increasing response priority.
[0066] Reinforcement learning is performed in a set Markov decision environment using deep learning algorithms. After receiving complaint data in the state space each time, the best response is selected in the action space. A reward function is set according to the governance objectives, based on resident satisfaction and resource utilization indicators. The reward function is adjusted according to resident feedback and resource utilization. Training stops when the model reaches the expected accuracy, and the complaint decision model is obtained.
[0067] When the complaint decision model receives a new complaint, it automatically generates a solution.
[0068] The state space includes the complaint type, priority, and processing progress, while the action space includes actions such as increasing manpower, extending working hours, and increasing response priority.
[0069] The specific process of modeling historical resident complaint data and corresponding solutions into a Markov decision process is as follows:
[0070] The state space is structured as follows: complaint categories are mapped to discrete state vectors using one-hot encoding, priorities are quantized into integer state values with a value range of 1-5 corresponding to [urgent, high, medium, low, normal], and processing progress is converted into continuous state values.
[0071] The action space is discretized as follows: adding manpower is subdivided into: {no addition, add 1 person, add 2 people}; extending working time is quantified into: {no extension, extend by 0.5 hours, extend by 1 hour}; and response priority is adjusted into: {maintain, upgrade by 1 level, upgrade by 2 levels}.
[0072] Construct a state transition probability matrix based on historical data, specifically by calculating the probability of transitioning from the current state to another state, as shown in the following formula:
[0073]
[0074] In the formula, P(s) ′ |s,a) indicates that the action a is performed to transition from the current state s to another state s. ′ The probability, N(s,a,s) ′ () indicates that action a is performed from the current state s to another state s. ′The historical count of the number of times action a is executed in the current state s; N(s,a) represents the historical count of the number of times action a is executed in the current state s.
[0075] Combine the state transition probabilities of each state into a state transition probability matrix;
[0076] Design a reward function.
[0077] Deep learning algorithms use either the Deep Q-Network (DQN) algorithm or the Proximal Policy Optimization (PPO) algorithm.
[0078] The specific process of reinforcement learning in a given Markov decision environment using the deep Q-network algorithm is as follows:
[0079] Construct an online Q-network and a target Q-network. The online Q-network uses a three-layer fully connected neural network; the target Q-network has the same structure as the online Q-network.
[0080] Initialize the experience replay pool D with samples (s, a, r, s'), where s represents the current state, a represents the action, r represents the reward, and s' represents the next state;
[0081] Obtain the current state from the state space, select the action corresponding to the state from the action space based on the ε-greedy policy; calculate the reward value of the current action and update the state; put the four-tuple of state, action, reward value and updated state into the experience replay pool;
[0082] Sample a batch of experience from the experience replay pool to update the Q value of the target Q network;
[0083] The loss function of the online Q-network is calculated using the updated Q-value of the target Q-network, and the parameters of the online Q-network are updated based on the gradient of the loss function; each In the next iteration, the parameters of the target Q network are updated by copying the parameters of the online Q network to the target Q network;
[0084] Once the model reaches the expected accuracy, training stops, and the complaint decision model is obtained; otherwise, training continues.
[0085] The specific process of reinforcement learning in a given Markov decision environment using the proximal policy optimization algorithm is as follows:
[0086] Construct an Actor-Critic network architecture, comprising an Actor network and a Critic network. States are collected from the state space and input into both the Actor and Critic networks respectively. The Actor network outputs the probability distribution of actions, and the Critic network outputs the estimated value of the states. Policy optimization employs a truncated objective function.
[0087] Calculate the reward value of the action, calculate the TD residual based on the estimated value and the reward value of the action, calculate the GAE advantage based on the TD residual, and construct the truncated objective function based on the GAE advantage and the probability distribution of the action.
[0088] The parameters of the Actor network and Critic network are continuously updated in the direction of gradient descent of the truncated objective function until the network converges, thus obtaining the complaint decision model.
[0089] The reward function is set according to the governance objectives as follows:
[0090] The reward function is designed as follows:
[0091] r = α·Sat + β·Util - γ·Cost
[0092] Where r represents the reward value, Sat represents the satisfaction reward, α is the weight of the satisfaction reward, Sat=2*exp(-0.5*t_delay)+0.3*rating, t_delay represents the portion of the actual processing time that exceeds the standard time, and rating represents the resident evaluation score;
[0093] Util represents the resource utilization reward, β is the weight of the resource utilization reward, Util = min(1, processed_tasks ÷ (staff_hours * 0.8)), processed_tasks represents the number of work orders processed per unit time, and staff_hours represents the total manpower hours invested;
[0094] Cost represents the penalty for resource consumption costs, and γ is the weight of the penalty for resource consumption costs. Cost = 0.2 * Δstaff + 0.1 * Δhours, where Δstaff represents the increased number of personnel and Δhours represents the extended working hours.
[0095] Example 2
[0096] This invention proposes a reinforcement learning-based complaint processing system corresponding to the method in Embodiment 1, comprising:
[0097] The data acquisition module is used to collect historical resident complaint data and corresponding solutions; the data acquisition module can be a smart terminal, sensor, smart camera, etc.
[0098] The model training module is used to train a complaint decision-making model based on collected historical resident complaint data and corresponding solutions;
[0099] The decision server is used to deploy the trained complaint decision model, run reinforcement learning algorithms, and generate decision recommendations by processing the collected data in real time.
[0100] The execution device is used to perform actual operations based on the decision recommendations of the decision server.
[0101] By using 5G networks or IoT protocols, the data acquisition module, model training module, decision server, and execution device are connected to ensure efficient data transmission and real-time processing.
[0102] The implementation methods of each module and its function in the system are completely consistent with the steps of the method in Implementation Example 1, so they will not be repeated here.
[0103] Example 3
[0104] The present invention proposes an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the complaint processing method based on reinforcement learning as described in Embodiment 1.
[0105] Example 4
[0106] The present invention proposes a computer-readable storage medium storing a computer program that causes a computer to execute the complaint processing method based on reinforcement learning as described in Embodiment 1.
[0107] In the embodiments disclosed in this application, a computer storage medium may be a tangible medium that may contain or store programs for use by or in conjunction with an instruction execution system, apparatus, or device. The computer storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatus, or devices, or any suitable combination of the foregoing. More specific examples of computer storage media include electrical connections based on one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the foregoing.
[0108] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.
[0109] The above are merely preferred embodiments of the present invention. The scope of protection of the present invention is not limited to the above embodiments. All technical solutions falling within the scope of the present invention's concept are within the scope of protection of the present invention. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of the present invention should be considered within the scope of protection of the present invention.
Claims
1. A complaint handling method based on reinforcement learning, characterized in that, Includes the following steps: Historical resident complaint data and corresponding solutions are collected and modeled as a Markov decision process. Based on the historical resident complaint data and corresponding solutions, a state space and an action space are defined. The state space includes the complaint type, priority, and processing progress, while the action space includes actions such as increasing manpower, extending working hours, and increasing response priority. Reinforcement learning is performed in a set Markov decision environment using deep learning algorithms. After receiving complaint data in the state space each time, the best response is selected in the action space. A reward function is set according to the governance objectives, based on resident satisfaction and resource utilization indicators. The reward function is adjusted according to resident feedback and resource utilization. Training stops when the model reaches the expected accuracy, and the complaint decision model is obtained. When the complaint decision model receives a new complaint, it automatically generates a solution.
2. The complaint handling method based on reinforcement learning as described in claim 1, characterized in that, The state space includes the complaint type, priority, and processing progress, and the action space includes actions such as increasing manpower, extending working hours, and increasing response priority. The specific process of modeling historical resident complaint data and corresponding solutions as a Markov decision process is as follows: The state space is structured as follows: complaint categories are mapped to discrete state vectors using one-hot encoding, priorities are quantized into integer state values with a value range of 1-5 corresponding to [urgent, high, medium, low, normal], and processing progress is converted into continuous state values. The action space is discretized as follows: adding manpower is subdivided into: {no addition, add 1 person, add 2 people}; extending working time is quantified into: {no extension, extend by 0.5 hours, extend by 1 hour}; and response priority is adjusted into: {maintain, upgrade by 1 level, upgrade by 2 levels}. Construct a state transition probability matrix based on historical data, specifically by calculating the probability of transitioning from the current state to another state, as shown in the following formula: In the formula, P(s) ′ |s,a) indicates that the action a is performed to transition from the current state s to another state s. ′ The probability, N(s,a,s) ′ () indicates that action a is performed from the current state s to another state s. ′ The historical count of the number of times action a is executed in the current state s; N(s,a) represents the historical count of the number of times action a is executed in the current state s. Combine the state transition probabilities of each state into a state transition probability matrix; Design a reward function.
3. The complaint handling method based on reinforcement learning as described in claim 1, characterized in that, The deep learning algorithm is a deep Q-network algorithm. The specific process of reinforcement learning in a given Markov decision environment using the deep Q-network algorithm is as follows: Construct an online Q-network and a target Q-network. The online Q-network uses a three-layer fully connected neural network; the target Q-network has the same structure as the online Q-network. Initialize the experience replay pool D, with samples (s, a, r, s'), where s represents the current state, a represents the action, r represents the reward, and s' represents the next state; Obtain the current state from the state space, select the action corresponding to the state from the action space based on the ε-greedy policy; calculate the reward value of the current action and update the state; put the four-tuple of state, action, reward value and updated state into the experience replay pool; Sample a batch of experience from the experience replay pool to update the Q value of the target Q network; The loss function of the online Q-network is calculated using the updated Q-value of the target Q-network, and the parameters of the online Q-network are updated based on the gradient of the loss function; each In the next iteration, the parameters of the target Q network are updated by copying the parameters of the online Q network to the target Q network; Once the model reaches the expected accuracy, training stops, and the complaint decision model is obtained; otherwise, training continues.
4. The complaint handling method based on reinforcement learning as described in claim 1, characterized in that, The deep learning algorithm is a proximal policy optimization algorithm. The specific process of performing reinforcement learning in a given Markov decision environment using the proximal policy optimization algorithm is as follows: Construct an Actor-Critic network architecture; This includes an Actor network and a Critic network. States are collected from the state space and input into the Actor and Critic networks respectively. The Actor network outputs the probability distribution of actions, and the Critic network outputs the estimated value of the states. Policy optimization uses a truncated objective function. Calculate the reward value of the action, calculate the TD residual based on the estimated value and the reward value of the action, calculate the GAE advantage based on the TD residual, and construct the truncated objective function based on the GAE advantage and the probability distribution of the action. The parameters of the Actor network and Critic network are continuously updated in the direction of gradient descent of the truncated objective function until the network converges, thus obtaining the complaint decision model.
5. The complaint handling method based on reinforcement learning as described in claim 1, characterized in that, The specific steps for setting the reward function based on the governance objectives are as follows: The reward function is designed as follows: r = α·Sat + β·Util - γ·Cost Where r represents the reward value, Sat represents the satisfaction reward, α is the weight of the satisfaction reward, Sat=2*exp(-0.5*t_delay)+0.3*rating, t_delay represents the portion of the actual processing time that exceeds the standard time, and rating represents the resident evaluation score; Util represents the resource utilization reward, β is the weight of the resource utilization reward, Util = min(1, processed_tasks ÷ (staff_hours * 0.8)), processed_tasks represents the number of work orders processed per unit time, and staff_hours represents the total manpower hours invested; Cost represents the penalty for resource consumption costs, and γ is the weight of the penalty for resource consumption costs. Cost = 0.2 * Δstaff + 0.1 * Δhours, where Δstaff represents the increased number of personnel and Δhours represents the extended working hours.
6. A complaint handling system based on reinforcement learning, characterized in that, include: The data collection module is used to collect historical resident complaint data and corresponding solutions; The model training module is used to train a complaint decision-making model based on collected historical resident complaint data and corresponding solutions; The decision server is used to deploy the trained complaint decision model, run reinforcement learning algorithms, and generate decision recommendations by processing the collected data in real time. The execution device is used to perform actual operations based on the decision recommendations of the decision server.
7. An electronic device, characterized in that, include: The system includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the reinforcement learning-based complaint processing method as described in any one of claims 1-6.
8. A computer-readable storage medium storing a computer program, characterized in that, The computer program causes the computer to perform the complaint handling method based on reinforcement learning as described in any one of claims 1-6.