ALE-Bench

ALE-Bench是一个评估人工智能在基于分数的算法编程比赛中表现的新基准，它收集了来自AtCoder Heuristic Contest (AHC)的真实任务，这些问题在计算上很难，并且没有已知的精确解决方案。ALE-Bench是一个长期基准，鼓励参与者迭代改进他们的解决方案，以适应没有已知最优解的优化问题。数据集包含40个AHC问题，涵盖了路由、规划、多智能体控制、解谜和贝叶斯推理等领域。这些问题是从AtCoder Inc.组织的AHC比赛中收集的，并通过Hugging Face平台发布，以确保数据的使用和授权。ALE-Bench旨在促进人工智能在算法工程领域的进步，特别是对于具有长期视野的问题解决能力。

ALE-Bench is a novel benchmark for evaluating AI performance in score-based algorithmic programming contests. It collects real-world tasks from AtCoder Heuristic Contest (AHC), which are computationally intractable and have no known exact solutions. As a long-term benchmark, ALE-Bench encourages participants to iteratively refine their solutions for optimization problems with no known optimal solutions. The dataset contains 40 AHC problems spanning domains such as routing, planning, multi-agent control, puzzle-solving, and Bayesian inference. These problems are collected from AHC contests organized by AtCoder Inc., and released via the Hugging Face platform to ensure proper data usage and licensing. ALE-Bench aims to advance AI progress in the field of algorithmic engineering, particularly for problem-solving capabilities with long-term planning horizons.