This repository provides the implementation for our paper:

📎 Paper Link: [arXiv/IEEE/DOI link]

Solving non-convex resource allocation problems poses significant challenges in wireless communication systems, often beyond the capability of traditional optimization techniques. To address this issue, we propose LLM-OptiRA, the first framework that leverages large language models (LLMs) to automatically detect and transform non-convex components into solvable forms, enabling fully automated resolution of non-convex resource allocation problems in wireless communication systems. LLM-OptiRA not only simplifies problem-solving by reducing reliance on expert knowledge, but also integrates error correction and feasibility validation mechanisms to ensure robustness. Experimental results show that LLM-OptiRA achieves an execution rate of 96% and a success rate of 80% on GPT-4, significantly outperforming baseline approaches in complex optimization tasks across diverse scenarios.

Figure 1: Overview of the LLM-OptiRA Framework

This figure demonstrates the LLM-OptiRA framework's process for optimizing wireless resource allocation using the GEO satellite user-beam allocation example, covering steps from problem description to result validation.

Clone the repository and install dependencies:

git clone https://github.com/Bearrr310/LLM-OptiRA
cd LLM-OptiRA
pip install -r requirements.txt

To run the main script:

linux:

export OPENAI_KEY=""

python main.py \
    --default_prompt_position Yourpath/OptiRA/Prompt/ \
    --default_dataset_position Yourpath/OptiRA/Dataset/ \
    --default_project_directory Yourpath/OptiRA/ \
    --api_key $OPENAI_KEY \
    --start_idx 0 \
    --end_idx 1 \
    --runs_per_problem 1

windows:

set OPENAI_KEY=

python main.py ^
    --default_prompt_position "Yourpath\LLM-OptiRA\Prompt" ^
    --default_dataset_position "Yourpath\LLM-OptiRA\Dataset" ^
    --default_project_directory "Yourpath\LLM-OptiRA" ^
    --api_key %OPENAI_KEY% ^
    --start_idx 0 ^
    --end_idx 1 ^
    --runs_per_problem 1

To run main.py, the following parameters should be provided:

• -default_prompt_position: The directory where the prompt templates are stored.

  • Example: Yourpath/OptiRA/Prompt/

• -default_dataset_position: The directory where the dataset files are stored.

  • Example: Yourpath/OptiRA/Dataset/

• -default_project_directory: The root directory of the project.

  • Example: Yourpath/OptiRA/

• -api_key: The OpenAI API key used for GPT-based computations.

  • Set this environment variable before running the script:

    export OPENAI_KEY="your-api-key"
    

• -start_idx: The starting index of the dataset problems to be processed.

• -end_idx: The ending index (exclusive) of the dataset problems to be processed.

  • The script processes problems in the range [start_idx, end_idx).
  • Example: -start_idx 1 --end_idx 3 will process problems indexed at 1 and 2, but not 3.

• -runs_per_problem: The number of times each problem should be executed.

  • Example: 5 (each problem in the selected range will be executed 5 times).

In main.py, the programs variable specifies which programs to run. You can execute the OptiRA approach and its baseline methods in parallel. By default, the code runs the following programs:

• OptiRA.py: Our OptiRA approach.
• baseline_Optimus.py: OptiMUS, a framework utilizing LLMs to formulate and solve Mixed Integer Linear Programming (MILP) problems from natural language descriptions.
• baseline_COT_refine.py: Chain-of-Thought (CoT), which generates a step-by-step reasoning process, encouraging the model to break down complex problems into smaller steps. It is enhanced with a refine stage from the LLM-OptiRA framework, which includes the ECL and FDC process.
• baseline_PS_refine.py: Plan-and-Solve (PS), which enhances LLMs' reasoning by dividing problem-solving into two stages: "planning" and "solving." This baseline method is paired with a refine stage from the LLM-OptiRA framework.
• baseline_COT.py: A basic version of Chain-of-Thought (CoT) without the refine stage.
• baseline_PS.py: A basic version of Plan-and-Solve (PS) without the refine stage.

To enable or disable specific baseline methods, you can modify the programs variable directly in main.py. For example, if you want to compare OptiRA with baseline_PS_refine.py, you can change it like this:

programs = [
    os.path.join(program_directory, 'OptiRA.py'),
    os.path.join(program_directory, 'baseline_PS_refine.py')
]

Figure 2: Comparison of Execution and Success Rates for Different Schemes using GPT-4 and GPT-3.5 Models

This figure shows that LLM-OptiRA outperforms baseline methods in execution and success rates on both GPT-4 and GPT-3.5 models, highlighting its effectiveness in solving non-convex problems.

Figure 3: Impact of Omitting Key Components of the LLM-OptiRA Framework on Success and Execution Rates across GPT4 and GPT-3.5

This figure reveals that removing components like convexification, ECL, or FDC significantly reduces the success and execution rates, emphasizing their importance in the framework.

Figure 4: Success and Execution Rates for LLM-OptiRA under Different Maximum Iteration Counts for ECL and FDC

This figure indicates that ECL and FDC improve success and execution rates, with ECL being effective in early iterations and FDC enhancing the success rate steadily as iterations increase.

For questions, please open an issue or reach out to bearrr310@outlook.com