ZeroFolio: Domain-Free Algorithm Selection

🔑 Enhanced Key Takeaways

• •ZeroFolio addresses the 'feature engineering bottleneck' in Algorithm Selection (AS) by bypassing the need for domain-specific feature extractors, which are notoriously difficult and expensive to design for new problem classes.
• •The approach leverages the inherent structural information present in raw problem files (e.g., DIMACS format for SAT), treating them as unstructured text to capture latent features via Large Language Model (LLM) embeddings.
• •By utilizing a non-parametric k-NN approach, ZeroFolio avoids the training overhead associated with deep learning-based end-to-end selectors, making it highly adaptable to new domains without retraining the core model.

🛠️ Technical Deep Dive

• •Embedding Strategy: Utilizes pretrained transformer-based encoders to map raw instance files into high-dimensional vector spaces.
• •Line Shuffling: A data augmentation technique applied to instance files to ensure the model remains invariant to the order of constraints or clauses, preventing overfitting to file formatting.
• •Distance Metric: Employs Manhattan distance (L1 norm) for k-NN, which has been empirically shown to be more robust than Euclidean distance in high-dimensional embedding spaces for this task.
• •Weighting Scheme: Implements inverse-distance weighting to prioritize the performance of the most similar historical instances when predicting the optimal algorithm for a new query.

🔮 Future ImplicationsAI analysis grounded in cited sources

⏳ Timeline