Skip to content

sums_and_differences_problems_1

Directory actions

More options

Directory actions

More options

Latest commit

 

History

History

sums_and_differences_problems_1

README.md

Sums and Differences (Ruzsa's Inequality) Example

This example demonstrates how to use the LoongFlow framework to solve a problem in additive combinatorics. The goal is to evolve a Python algorithm that constructs a set of integers $A$ with specific sumset and difference set properties.

Problem Description

The objective is to find a set of integers $A$ (where $|A| = n$) that maximizes the ratio related to Ruzsa's inequality. Specifically, we want to maximize the score derived from the sizes of the sumset $A+A$ and the difference set $A-A$.

The score is calculated as: $$ \text{Score} = \frac{\ln(|A+A|/|A|)}{\ln(|A-A|/|A|)} $$

Where:

  • $A+A = {a+b \mid a, b \in A}$
  • $A-A = {a-b \mid a, b \in A}$

For the detailed problem context (Note: Link provided points to the repository source): Problem Source Link

In this specific configuration, the system evolves a search algorithm to find the optimal list of integers.

Project Structure

  • initial_program.py: The starting seed code. It contains a stochastic search function search_for_best_set that mutates a list of integers to improve the score.
  • eval_program.py: The evaluation logic. It executes the generated code, validates the output format (Tuple[np.ndarray, str]), ensures elements are integers, and calculates the score using Numba-optimized functions.
  • task_config.yaml: The main configuration file defining the LLM prompt (DeepSeek), evolution parameters (iterations, target score), and the agent components (Planner, Executor, Summarizer).

How to Run

To start the evolution process, you need to use the math_agent_agent.py entry point. Ensure your PYTHONPATH includes the project root so that python can find the agents and evolux modules.

1. Prerequisites

Ensure you are in the root directory of your local project (the directory containing agents/ and evolux/).

2. Execution Command

Run the following command to kick off the evolution. This command loads the base configuration and injects the initial code and evaluation logic from the respective files.

python agents/math_agent/math_agent_agent.py \
  --config agents/math_agent/examples/sums_and_differences_problems_1/task_config.yaml \
  --initial-file agents/math_agent/examples/sums_and_differences_problems_1/initial_program.py \
  --eval-file agents/math_agent/examples/sums_and_differences_problems_1/eval_program.py \
  --log-level INFO

Arguments Explanation:

  • --config: Path to the YAML configuration file (task_config.yaml).
  • --initial-file: Path to the Python file containing the seed code (initial_program.py). The content of this file will be injected into evolve.initial_code.
  • --eval-file: Path to the Python file containing the evaluation logic (eval_program.py). The content will be injected into evolve.evaluator.evaluate_code.
  • --log-level: Sets the logging verbosity (e.g., INFO, DEBUG).

3. Configuration Highlights

The task_config.yaml is pre-configured with the following strategies:

  • Planner: evolve_planner (Handles the strategic direction of code modification).
  • Executor: evolve_executor_fuse (A powerful executor that fuses multiple thought processes/candidates, configured with a high score threshold).
  • Summarizer: evolve_summary (Summarizes the results of the execution for the next iteration).
  • LLM: Configured to use deepseek-r1-250528.

Evolution Process & Results

The system iterates through generations of code, attempting to maximize the ratio described above.

Final Result

The best solution found by LoongFlow in the latest run achieved a significant improvement over the baseline.

Result Metrics:

  • Best Score Achieved: 1.103534711409646
  • Previous Baseline: 1.059793

Troubleshooting

  • TimeoutError: The eval_program.py enforces a strict timeout (default 3600s in config, though internal function calls have shorter timeouts). If the generated code enters an infinite loop, it will be terminated and marked as a failure.
  • ModuleNotFoundError: Ensure your PYTHONPATH is set correctly. You may need to run export PYTHONPATH=$PYTHONPATH:. in the project root before running the command.
  • Numba Compilation Overhead: The first execution might be slightly slower due to JIT compilation of the get_score_numba function.