EliteTable: Self-Improving Program Search for Executable Control Policies

查看原文 / View Original 编辑 / Edit

Computer Science > Artificial Intelligence arXiv:2610.12536 (cs) [Submitted on 8 Oct 2026] Title:EliteTable: Self-Improving Program Search for Executable Control Policies Authors:Stuart Farmer View a PDF of the paper titled EliteTable: Self-Improving Program Search for Executable Control Policies, by Stuart Farmer View PDF HTML (experimental)

计算机科学 > 人工智能 arXiv:2610.12536 (cs) [提交于 2026 年 10 月 8 日] 标题：EliteTable：用于可执行控制策略的自我改进程序搜索作者：Stuart Farmer 查看题为《EliteTable：用于可执行控制策略的自我改进程序搜索》的论文 PDF，作者：Stuart Farmer 查看 PDF HTML（实验性）

Abstract:EliteTable is a self-improving program-search algorithm guided by large language models (LLMs) that uses a fixed-capacity table of high-scoring solutions to guide new proposals, edits of existing solutions, and remixes of several solutions. Performance was evaluated across six Gymnasium control tasks, each with ten generations and fifty candidate slots per generation. Each task was evaluated across ten independently initialized searches.

摘要：EliteTable 是一种由大语言模型 (LLM) 指导的自我改进程序搜索算法，它使用一个固定容量的高分解决方案表来引导新的提案、对现有解决方案的编辑以及对多个解决方案的重组。性能在六项 Gymnasium 控制任务中进行了评估，每项任务包含十代，每代有五十个候选槽位。每项任务均通过十次独立初始化的搜索进行了评估。

Candidate-score medians rose between the first and final generations in all six tasks, and selected-winner search-score medians rose in five. In CartPole-v1, the selected winner reached the search-score ceiling in the first generation of all ten searches. All ten paired held-out endpoints improved in four of the six tasks. BipedalWalker-v3 and CarRacing-v3 improved without reaching their reporting targets.

在所有六项任务中，候选分数中位数在第一代和最后一代之间均有所上升，且选定获胜者的搜索分数中位数在五项任务中有所上升。在 CartPole-v1 中，选定的获胜者在所有十次搜索的第一代中均达到了搜索分数的上限。在六项任务中的四项里，所有十个配对的留出端点均有所改善。BipedalWalker-v3 和 CarRacing-v3 虽有改善，但未达到其报告目标。

Six of ten Acrobot-v1 searches declined on held-out episodes, and the median paired change was negative. The results document improvement during program search and the transfer of selected-winner gains to new episodes in some tasks. Comparisons and ablations are needed to establish EliteTable’s relative performance and the contributions of its components.

在十次 Acrobot-v1 搜索中，有六次在留出剧集中表现下降，且配对变化的中位数为负值。结果记录了程序搜索过程中的改进，以及在某些任务中将选定获胜者的收益转移到新剧集的情况。需要进行比较和消融研究，以确定 EliteTable 的相对性能及其各组件的贡献。