Scholarlink

<p>Evolutionary algorithms are widely used for the resource-constrained project scheduling problem (RCPSP), but their performance often depends on hand-crafted priority rules, repair operators, and constraint-handling strategies. This paper presents VLCEA, a validated framework that uses a large language model (LLM) not as a direct solver but as a source of instance-specific guidance for evolutionary scheduling. The LLM generates activity-level priority scores, whereas ordering hints are constructed deterministically from direct precedence relations already present in the instance, thereby preventing hallucinated constraints by design. To ensure robustness and reproducibility, generated outputs are subjected to JSON Schema validation and semantic checks, with bounded retries and a deterministic fallback when generation fails. The validated guidance is integrated into a permutation-based evolutionary algorithm with an SSGS decoder in three ways: chromosome repair, decoder bias, and fitness-penalty shaping. Experiments on PSPLIB J30 and larger PSPLIB sets (J60, J90, and J120), evaluated against the reference values available for each benchmark class, show a consistent pattern: repair-based injection is the most effective way to use the generated guidance, achieving performance close to the hand-crafted repair baseline across all tested instance classes, whereas decoder-only bias yields limited gains and fitness-penalty shaping often degrades performance. On the largest tested set, J120, the constraint-generation pipeline validated all 600 instances with a deterministic fallback rate of 1.33%, suggesting that the proposed validation and fallback protocol remains operational at larger scale. These results indicate that validated, instance-grounded LLM guidance can improve evolutionary RCPSP scheduling when injected at appropriate points in the search pipeline under the tested fixed-budget setting.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

VLCEA: LLM-guided evolutionary scheduling for RCPSP via validated, hallucination-safe constraints

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年8月4日

查看全文
  • Hyeonwoo Kim,
  • Wonho Song,
  • Hyungjoon Kim

摘要

Evolutionary algorithms are widely used for the resource-constrained project scheduling problem (RCPSP), but their performance often depends on hand-crafted priority rules, repair operators, and constraint-handling strategies. This paper presents VLCEA, a validated framework that uses a large language model (LLM) not as a direct solver but as a source of instance-specific guidance for evolutionary scheduling. The LLM generates activity-level priority scores, whereas ordering hints are constructed deterministically from direct precedence relations already present in the instance, thereby preventing hallucinated constraints by design. To ensure robustness and reproducibility, generated outputs are subjected to JSON Schema validation and semantic checks, with bounded retries and a deterministic fallback when generation fails. The validated guidance is integrated into a permutation-based evolutionary algorithm with an SSGS decoder in three ways: chromosome repair, decoder bias, and fitness-penalty shaping. Experiments on PSPLIB J30 and larger PSPLIB sets (J60, J90, and J120), evaluated against the reference values available for each benchmark class, show a consistent pattern: repair-based injection is the most effective way to use the generated guidance, achieving performance close to the hand-crafted repair baseline across all tested instance classes, whereas decoder-only bias yields limited gains and fitness-penalty shaping often degrades performance. On the largest tested set, J120, the constraint-generation pipeline validated all 600 instances with a deterministic fallback rate of 1.33%, suggesting that the proposed validation and fallback protocol remains operational at larger scale. These results indicate that validated, instance-grounded LLM guidance can improve evolutionary RCPSP scheduling when injected at appropriate points in the search pipeline under the tested fixed-budget setting.