arXivDaily arXiv每日学术速递 周一至周五更新
北美退伍军人节(Veterans Day),12 日无更新
arXiv 2602.22260cs.LGcs.NE
收藏

用于进化算法参数控制的代码世界模型

Code World Models for Parameter Control in Evolutionary Algorithms

ICN2 -- Catalan Institute of Nanoscience;Nanotechnology, Barcelona, Spain Artificial Intelligence Research Institute (IIIA-CSIC), Barcelona, Spain Universitat Polit\` e cnica de Catalunya (UPC), Barcelona, Spain

Camilo Chacón Sartori, Guillem Rodríguez Corominas

AI总结:

本文提出利用代码世界模型进行进化算法参数控制,通过模拟器实现高效突变强度选择,在多个基准测试中表现出色。

摘要:

可以一个大语言模型学习优化器的行为,并利用这种知识来控制它吗?我们扩展了代码世界模型(CWMs),即由大语言模型生成的Python程序,这些程序能够预测环境动态,从确定性游戏扩展到随机组合优化。给定(1+1)-RLS_k的次优轨迹,大语言模型会生成一个优化器动态的模拟器;在该模拟器上进行贪婪规划,然后在每一步选择突变强度k。在lo和onemax上,CWM-贪婪的表现接近理论最优策略——在从未见过最优策略轨迹的情况下。在jump_k上,一个欺骗性山谷导致所有自适应基线失败(0%的成功率),CWM-贪婪实现了100%的成功率——没有任何使用oracle知识的策略。在NK-景观上,其中不存在闭式模型,当提示包含经验转移统计时,CWM-贪婪在十五个独立生成的实例中优于所有基线(36.94 vs. 36.32;p<0.001)。CWM在样本效率(200条离线轨迹 vs. 500条在线回合)、成功率(100% vs. 58%)和泛化(k=3:78% vs. 0%)方面均优于DQN。鲁棒性实验确认了在5次独立运行中的稳定合成。

英文摘要:

Can an LLM learn how an optimizer behaves -- and use that knowledge to control it? We extend Code World Models (CWMs), LLM-synthesized Python programs that predict environment dynamics, from deterministic games to stochastic combinatorial optimization. Given suboptimal trajectories of $(1{+}1)$-$\text{RLS}_k$, the LLM synthesizes a simulator of the optimizer's dynamics; greedy planning over this simulator then selects the mutation strength $k$ at each step. On \lo{} and \onemax{}, CWM-greedy performs within 6\% of the theoretically optimal policy -- without ever seeing optimal-policy trajectories. On \jump{$_k$}, where a deceptive valley causes all adaptive baselines to fail (0\% success rate), CWM-greedy achieves 100\% success rate -- without any collection policy using oracle knowledge of the gap parameter. On the NK-Landscape, where no closed-form model exists, CWM-greedy outperforms all baselines across fifteen independently generated instances ($36.94$ vs.\ $36.32$; $p<0.001$) when the prompt includes empirical transition statistics. The CWM also outperforms DQN in sample efficiency (200 offline trajectories vs.\ 500 online episodes), success rate (100\% vs.\ 58\%), and generalization ($k{=}3$: 78\% vs.\ 0\%). Robustness experiments confirm stable synthesis across 5 independent runs.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑