Large Language Model-Enhanced Algorithm Selection: Towards Comprehensive Algorithm Representation
Authors: Xingyu Wu, Yan Zhong, Jibin Wu, Bingbing Jiang, Kay Chen Tan
IJCAI 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Extensive experiments validate the performance superiority of the proposed model and the efficacy of each key module. |
| Researcher Affiliation | Academia | Xingyu Wu1 , Yan Zhong2 , Jibin Wu1 , Bingbing Jiang3 and Kay Chen Tan1 1The Hong Kong Polytechnic University 2Peking University 3Hangzhou Normal University |
| Pseudocode | No | No pseudocode or algorithm blocks explicitly presented. |
| Open Source Code | Yes | The implementation of AS-LLM is available at https://github.com/wuxingyu-ai/AS-LLM |
| Open Datasets | Yes | ASlib (Algorithm Selection Library) Benchmark2 [Bischl et al., 2016] is a standardized dataset for algorithm selection problems, aimed at providing a common benchmark for evaluating and comparing the performance of different algorithm selection methods. 2https://www.coseal.net/aslib/ |
| Dataset Splits | No | Prior to each experiment, we randomly select 80% of the problems as the training set, while the remaining 20% constitute the test set. |
| Hardware Specification | No | No specific hardware details (e.g., GPU/CPU models, memory) were provided for the experimental setup. |
| Software Dependencies | No | While AS-LLM uniquely incorporates algorithm code or relevant text, which are processed by the respective LLMs (Uni XCoder [Guo et al., 2022] and BGE [Xiao et al., 2023]) to extract algorithm features, no specific version numbers for these or other software dependencies are provided. |
| Experiment Setup | Yes | This includes adjusting the number of layers in each MLP, choosing suitable activation functions, determining the number of neurons in each hidden layer of the components, adjusting the parameters in feature selection module, and determining the incorporation of the MLP module after similarity calculation. Additionally, we explore the fusion of two distinct algorithm representation vectors, i.e., the LLM representation and embedding layer representation, and assign pre-defined weights α in Eq. (6) before conducting the experiments. |