【作者】 王宁；

【导师】 张强；

【作者基本信息】 大连理工大学 ， 计算机科学与技术， 2025， 硕士

【摘要】 随着虚拟现实、仿真训练与电子游戏等场景生成应用需求的持续增长,如何高效地在复杂设计空间中生成高质量、具备多样性与适应性的场景内容成为研究热点。传统的程序内容生成方法虽在精度与控制性上具备优势,但面对复杂多变的用户需求与设计环境时,往往在灵活性与扩展性上显得不足。进化算法因其强大的搜索能力与良好的自适应特性,在场景生成任务中被广泛应用,成为提升内容生成效率与质量的重要手段。然而,进化算法在适应度函数设计与多样性维护方面仍面临诸多挑战。质量多样性算法应运而生,致力于在行为空间中寻找性能优异且风格多样的解集。基于此背景,本文考虑开展面向质量多样性的场景内容生成优化方法研究,融合多模态大语言模型评估能力与自适应兴趣驱动的多发射器搜索策略,旨在提升生成内容的评价效率、优化搜索策略的灵活性与鲁棒性,从而在内容质量与多样性之间实现良好平衡。针对场景内容生成中进化算法在适应度函数设计上所面临的挑战,本文提出了一种基于多模态大模型适应度评估的进化内容生成方法,创新性地将大语言模型引入进化算法的适应度评估环节,用于辅助进化过程对生成内容质量与多样性的判断。通过构建多维度提示模板与内容对比策略,该方法充分利用大语言模型在语言理解与图文处理方面的优势,为每个候选解提供多角度、多模态的质量反馈。为进一步提升评估效率,本文还构建了一个轻量级蒸馏模型对大语言模型评估结果进行近似学习,有效降低计算成本。最后,在经典的马里奥游戏关卡生成任务中进行了验证。结果表明,该框架不仅能显著提升生成内容的视觉表现力与结构合理性,还在用户主观评价中获得了高度认可,展现出优异的泛化能力与实用性,为复杂生成任务提供了全新的适应度评估范式。针对任务需求持续变化及内容元素不断更新所带来的挑战,本文提出了一种面向动态场景优化的自适应兴趣驱动多发射器优化方法,增强了传统质量多样性算法在动态环境中的适应能力。该方法由兴趣发射器构建模块与搜索空间探索模块组成,前者可围绕特定游戏元素构建专用发射器,后者借鉴协方差矩阵自适应进化策略的重启机制,实现搜索方向与强度的自适应调控。同时,框架允许对每个解的存储单元设置动态接受阈值,以精准控制发射器行为,有效规避行为维度爆炸带来的搜索失效问题。在马里奥关卡生成实验中,该方法展现出对新元素快速适应与高效探索的能力,并且支持设计者从内容生成过程中获取设计启发。更多还原

【Abstract】 With the growing demand for scene generation in virtual reality,simulation training,and video games,how to efficiently generate high-quality,diverse,and adaptive content in a complex design space has become a hot research topic.Although traditional procedural content generation methods have advantages in accuracy and controllability,they often lack flexibility and scalability when facing changing user needs and complex design environments.Due to their strong search ability and adaptive features,evolutionary algorithms have been widely used in scene generation tasks and have become an important tool to improve the efficiency and quality of content generation.However,challenges still remain in designing fitness functions and maintaining diversity.To solve this,quality-diversity algorithms have emerged,aiming to find a set of solutions with both good performance and diverse styles in the behavior space.Based on this background,this paper explores a new optimization method for scene content generation under the framework of quality-diversity algorithms.It combines the evaluation ability of multimodal large language models with an adaptive,interest-driven multi-emitter search strategy.The goal is to improve evaluation efficiency and enhance the flexibility and robustness of the search process,achieving a better balance between content quality and diversity.To address the difficulty of fitness function design in evolutionary content generation,we propose an approach that introduces multimodal large models into the fitness evaluation stage.This method uses large language models to assist in judging the quality and diversity of generated content.By designing multi-dimensional prompt templates and comparison strategies,the method makes full use of the model’s advantages in language understanding and image-text processing,providing multi-angle,multimodal feedback for each candidate solution.Furthermore,we build a lightweight distillation model to learn the evaluation results of the large model,which helps reduce computation cost while maintaining accuracy.Finally,we verify the method in a classic Super Mario level generation task.The results show that our framework significantly improves the visual quality and structural rationality of generated content.It also receives high ratings in user evaluations,demonstrating excellent generalization and practical value,and offers a new paradigm for fitness evaluation in complex generation tasks.To deal with changing task requirements and the continuous update of content elements,we further propose an adaptive interest-driven multi-emitter optimization method for dynamic scene generation.This method enhances the adaptability of traditional quality-diversity algorithms in dynamic environments.It includes two key modules:the emitter construction module and the search exploration module.The former allows for designing specific emitters for new game elements,while the latter adopts a restart strategy from covariance matrix adaptation evolution strategy to control the search direction and strength adaptively.In addition,the framework supports setting dynamic acceptance thresholds for each solution container,allowing fine control of emitter behavior and avoiding search failure caused by high-dimensional behavior space.Experiments in Mario level generation show that this method can quickly adapt to new elements,support efficient exploration,and provide useful design insights for content creators.更多还原