Table of contents
Answering process for question 1:
Answering process for question 3:
Decision optimization application scenarios: artificial intelligence model hyperparameter tuning
To obtain the complete code paper, see this business card
(CPN) is a new type ofinformation infrastructure. See the complete paper code at the end of the article
Allocate and schedule computing resources according to business needs, usually consisting ofend users, edge servers and cloud servers. The network is designed to meet the needs of a variety of computing tasks. Strategically allocating computing resources based on the spatial distribution of computing needs helps reduce latency, lower costs, and improve overall network efficiency and user experience. Utilizingoperations researchand optimization techniques to model the location selection and layout of information infrastructure can help make more scientific decisions from a global optimization perspective . This approach helps improve decision-making efficiency and planning effectiveness. As the size of the computational challenge grows, the computational complexity of accurately solvingcombinatorial optimizationproblems increases exponentially. Existing workarounds often struggle to complete these solutions in a reasonable amount of time. In addition, as traditional computing power approaches the limits defined by Moore's Law, quantum computing is expected to break through the traditional computing power. Computercomputation bottleneck technology emerged. Coherent Easy Camera (CIM), as a dedicated quantum computing device, introduces a new perspective in efficiently solving operation optimization problems in various industries. Due to its ability to solve It has received widespread attention due to its excellent performance in combinatorial optimization problems. Due to its close relationship with the easy-phase model, the QUBO (Quadratic UnconstrainedBinaryOptimization) model forms a core problem class in quantum computing, which can be Solved using Coherence Easy Camera (CIM). In this competition, the main focus is on optimizing the layout of computing network infrastructure. The problem is modeled using QUBO form, and the solution is implemented using Kaiwu SDK. Kaiwu SDK is a software development kit designed for solving QUBO models using Coherent Easy Camera (CIM). You can obtain the necessary resources by visiting this link [1] and clicking on the "Download" option.
Problem Description
This competition involves the layoutoptimization of computing network infrastructure in a specific area. The area is divided into several adjacentsquaregrids, and the compute demand distribution data provides the aggregate compute demand within each grid. The coordinates in the data represent the center coordinates of each grid. To simplify the problem, the computational demand points within each grid are merged into the center point of the grid (i.e., each grid is considered to correspond to a single demand point). The computing requirements within the grid are generated by end devices connected to the network, such as sensors, smartphones, industrial robots, etc. Computing needs within the computing network are met by edge servers andcloud servers. Edge servers are located at the "edge" of the network, usually close to users or devices. Their role is to process data closer to the user, thereby increasing responsiveness and efficiency. Because edge servers are closer to users, they can handle requests more quickly, offloading the core cloud infrastructure and improving overall operational efficiency. On the other hand, cloud servers are located in data centers far away from users and have powerful computing and storage capabilities. When the capacity of edge servers is insufficient, cloud servers can serve as supplementary resources. The collaborative interaction between edge servers and cloud servers optimizes the performance and reliability of the entire system.
2.1 Question 1
The task is to deploy two edge servers in the grid area according to the computing requirements. The coverage radius of each edge server is 1. For example, in Figure 2, we present a schematic diagram of the coverage effect of an edge server with a coverage radius of 2. The goal is to determine the location of two edge servers to cover the greatest computing needs. Assuming that the location of the edge server is in the center of the grid, the attachment (Attachment 1_Computational Demand Distribution Data.csv) provides the computational demand within each grid. Please formulate a QUBO model for the problem and use Kaiwu SDK's Simulated Annealing solver and CIM simulator to solve it. Provides the coordinates of edge servers deployed covering the largest computing needs, and the corresponding total computing demand coverage. Figure 2: Schematic diagram of an edge server with a coverage radius of 2, with coordinates [3,3]. (About coverage determination: If the center distance of the grid (Euclideandistance) is less than or equal to the coverage radius of the edge server, it is considered to be This edge server covers.)
2.2 Question 2
When the edge server cannot meet the computing demand, the computing service will be provided by the upstream cloud server. Edge servers and end users can choose to connect to cloud servers. When the computing demand received by the edge server exceeds its capacity limit, the edge server allocates the excess demand of the edge server directly to the cloud server. Each end user's needs must be met and can only connect to one server, either a cloud server or an edge server. The computing server has a resource capacity limit, assuming that the available computing resource capacity of each edge server is 12, while the cloud server has unlimited computing resource capacity (ignoring any resource limitations of the cloud server). The server also has a coverage radius; assuming that the edge server's coverage radius is 3, and the cloud server's coverage radius is infinity. Deploying edge servers typically incurs costs, which include fixed costs, compute costs, and transport costs. Fixed costs only depend on whether edge servers are deployed at the candidate locations. The compute cost is proportional to the amount of compute resources requested and is calculated as the unit compute cost multiplied by the compute load. The unit computing cost of cloud servers is 1 and that of edge servers is 2. In addition, the transmission cost exists in the transmission between user end to edge, edge to cloud, and user end to cloud, and is calculated by multiplying the computing demand times the transmission distance times the unit transmission cost. The Euclidean distance of the transmission distance is rounded to two decimal places and calculated as the one-way distance (ignoring round-trip transmission). The unit transmission cost from the user to the edge and from the edge to the cloud is 1, and the unit transmission cost from the user to the cloud is 2. Please provide a computing network layout that meets the computing needs of all clients in the region at the minimum total cost. This includes the location and number of edge servers, as well as user-to-edge, edge-to-cloud, and user-to-cloud server connections. Formulate a QUBO model using as few bits as possible (the SDK only supports problem instances up to 100 bits) and solve it using Kaiwu SDK's simulated annealing and CIM simulator for solving. The computing needs of each user, corresponding to their geographical location, are provided in the user data file (Attachment 2_Computational Demand Distribution Data.csv). The edge data file (Attachment 3_Candidate Edge Facilities Data.csv) provides candidate coordinates for deploying edge servers and the fixed costs associated with each location. Cloud data file (Attachment 4_Cloud Facilities Data.csv) provides the coordinates of the cloud server.
2.3 Question 3
Please propose a possible decision optimization application scenario that can build a suitable QUBO model. Suggested scenario areas include, but are not limited to, artificial intelligence, big data, cloud computing andedge computing. The scenario should have real value, scalability, real business needs, and demonstrate the benefits of Coherent Easy Camera (CIM). Provide necessary background information, research methodology, methods, expected research results, technology roadmap and supporting references or materials. Report length should be between 500 and 1000 words.
Answering process for question 1:
Answering process for question 2:
Answering process for question 3:
Decision Optimization Application Scenario: Artificial Intelligence ModelHyperparameter Tuning
Background Information:
In the field of artificial intelligence, the performance of machine learning models often depends on the selection and adjustment of hyperparameters. Hyperparameters are parameters set before training the model. Different combinations of hyperparameters can significantly affect the performance of the model. Traditional hyperparameter tuning methods often require multiple experiments, consuming a lot of time and computing resources. Therefore, leveraging the advantages of quantum computing to accelerate the hyperparameter tuning process of artificial intelligence models has become a promising research task.
In current deep learning and machine learning research, researchers often need to face complex model structures and large-scale data sets, which makes the search space for hyperparameters huge. Reasonable selection of hyperparameters can significantly improve the performance of the model, but it also increases the difficulty of hyperparameter tuning. Traditional methods use greedy search or grid search, etc., but these methods operate in high-dimensional space The efficiency is lower in , especially when there are complex interrelationships between hyperparameters.
Research methods:
1. Problem definition: We hope to use the advantages of quantum computing to accelerate the hyperparameter tuning process of artificial intelligence models. Specifically, we treat hyperparameters as decision variables and the model's performance measure (e.g., accuracy) as the objective function.
2. QUBO model construction: We introduce each hyperparameter into a binary variable to build a QUBO model. The objective function aims to maximize or minimize a performance measure, taking into account the interrelationships between hyperparameters.
3. Applications of Coherent Ising Machines (CIM): Utilizing CIM’s parallel computing and high degree of connectivity to be more efficient in quantum computing Search the superparameter space. The quantum advantage of CIM can improve search efficiency and find better hyperparameter combinations.
Scenario advantages:
1. Efficient search: The parallelism and highly connected nature of quantum computing enable more efficient searches for optimal solutions in high-dimensional, complex hyperparameter spaces.
2. Accelerate model training: By finding the optimal hyperparameter combination more quickly, the time for model training can be reduced and the iteration speed of the artificial intelligence model can be improved.
3. Widely applicable: This scenario is not only applicable to deep learning models, but also to other machine learning algorithms, expanding the application field.
Expected research results:
By introducing quantum computing into the hyperparameter tuning process of artificial intelligence models, we expect to achieve the following goals:
1. Faster model tuning process to improve the performance of artificial intelligence models.
2. Explore the advantages of quantum computing in hyperparameter optimization and lay the foundation for wider applications in the field of artificial intelligence in the future.
3. Provide a practical method that enables researchers to adjust hyperparameters more efficiently and promote the progress of artificial intelligence research.
Technical route:
1. Select key hyperparameters:
First, we need to carefully select key hyperparameters that have a greater impact on model performance. This may include learning rate, number of layers, number of nodes, etc., depending on the machine learning model and task used.
2. QUBO model design:
2.1 Introduction of decision variables:
For each selected hyperparameter, a binary variable is introduced to represent its value. For example, for the hyperparameter \(p_i\), introduce binary variables
2.2 Design of objective function:
Design the objective function to target the performance measure of the model and try to maximize or minimize this performance measure. The form of the objective function isquadratic
3. CIM solution:
Use Coherent Ising Machines (CIM) to solve the built QUBO model. CIM takes full advantage of quantum computing and can efficiently solve high-dimensional and complex problems.
4. Performance evaluation:
Train a machine learning model using the optimal combination of hyperparameters obtained and evaluate its performance on the validation or test set. Compare the performance of hyperparameter combinations obtained using quantum computing methods and traditional methods.
Model building process:
1. Select a machine learning model: Determine the machine learning model to be tuned for hyperparameters, such asdeep neural network, support vector machine wait.
2. Select key hyperparameters: Carefully select key hyperparameters that have a greater impact on model performance. This may require the experience of domain experts or be determined through experimental analysis.
3. QUBO model construction:
- Introduce decision variables: Introduce a binary variable for each hyperparameter.
- Objective function design: Taking the performance measurement of the model as the goal, design an objective function that can be solved in quantum computing.
- Introduction of constraints: Introduce constraints for hyperparameters to ensure that their values are within a reasonable range.
4. CIM solution: Use Coherent Ising Machines to solve the constructed QUBO model to obtain the optimal hyperparameter combination.
5. Performance evaluation:
- Train machine learning models using optimal hyperparameter combinations.
- Evaluate the performance of the model on the validation or test set.
- Compare the performance of hyperparameter combinations obtained using quantum computing methods and traditional methods.
Through this technical route, we expect to give full play to the advantages of quantum computing in hyperparameter tuning of artificial intelligence models, improve the efficiency of the optimization process, and better support research and applications in the field of artificial intelligence.