서강대학교

목차

List of Figures iv
List of Tables v
초록 vi
Abstract viii
I. Introduction 1
1.1 Background and Motivation 1
1.2 Problem Statement 2
1.3 Contributions 3
1.4 Thesis Organization 3
II. Related Work 5
2.1 Symbolic Regression and Explainable Modeling 5
2.2 Neural and LLM-guided Equation Discovery 6
2.3 Visual and Multimodal Symbolic Regression 6
2.4 LLM-based Evolutionary Program Search 7
2.5 Chart Understanding and Visual Grounding 8
2.6 VLM Backbones 8
2.7 Image-space Fitness and Visual Curve Matching 9
III.EvoVLM 10
3.1 Method Overview 10
3.2 Problem Formulation and Notation 11
3.3 Adaptive RBF-PELT Segmentation 12
3.4 Multimodal Segment Context 16
3.4.1 Primary Graph Rendering 16
3.4.2 Auxiliary Visual Evidence 16
3.4.3 Textual Statistical Context 18
3.5 Topology-Aware Formula Prior Retrieval 20
3.6 Archive-Guided Evolutionary Search 20
3.6.1 Candidate Formula Parameterization 21
3.6.2 Piecewise Structural Formula Components 22
3.6.3 Active Selection Objectives 23
3.6.4 Archive Feedback Mechanism 25
3.7 Algorithm Summary 26
3.8 Post-hoc Structural Diagnostics 26
IV.Experiments 30
4.1 Experimental Setup 30
4.1.1 Datasets 30
4.1.2 Implementation Details and Evaluation Protocol 31
4.2 Overall Performance and Baseline Comparison 32
4.3 Structural Analysis of Discovered Expressions 33
4.3.1 PySR-best Flight Delay Inspection 34
4.3.2 Dataset-wise EvoVLM-R2 Representative Comparisons 35
4.4 Prompt Configuration 38
4.4.1 EvoVLM-R2 38
4.4.2 EvoVLM-IoU 39
4.4.3 Cross-track Prompt Sensitivity 40
4.5 Auxiliary View Analysis 41
4.5.1 EvoVLM-R2 41
4.5.2 EvoVLM-IoU 42
4.6 Generation Dynamics 43
4.7 Backbone Comparison 44
4.8 Computational Efficiency 46
V. Discussion 47
5.1 Interpretation of Findings 47
5.2 Limitations and Future Directions 48
VI.Conclusion 50
References 52