Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Sequential Multi-Agent Dynamic Algorithm Configuration
Authors: Chen Lu, Ke Xue, Lei Yuan, Yao Wang, Yaoyuan Wang, Sheng Fu, Chao Qian
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Experiments from synthetic functions to the configuration of multi-objective optimization algorithms demonstrate Seq-MADAC’s superior performance over state-of-the-art MARL methods and show strong generalization across problem classes. |
| Researcher Affiliation | Collaboration | 1 National Key Laboratory for Novel Software Technology, Nanjing University, China 2 School of Artificial Intelligence, Nanjing University, China 3 Advanced Computing and Storage Lab, Huawei Technologies Co., Ltd. Shenzhen, China |
| Pseudocode | Yes | Algorithm 1 Benchmark Outline: Seq-Sigmoid |
| Open Source Code | Yes | Our code is available at https://github.com/lamda-bbo/seq-madac. |
| Open Datasets | Yes | The problem instances of MOEA/D include the well-known multi-objective optimization problems (MOP) benchmarks DTLZ [8] and WFG [13] with variable numbers of objectives and problem dimensions, which cover different difficulty levels of MOPs. |
| Dataset Splits | Yes | On the original Sigmoid benchmark, we traditionally train the agents with a set of given instances (i.e., a set of si,h and pi,h) and test them on other sampled instances. In order to better reflect the generalization ability of the algorithms, we resample the instance (i.e., resample a new si,h and pi,h) every time one episode is done and the environment is reset. ... As demonstrated in Table 1, the top three problems are the problems for training, and the bottom five problems are for testing. |
| Hardware Specification | No | All the experiments are easy to carry out, which has little requirement for computer resources, and the detailed information of the computer resources we use see Appendix D. (Appendix D not provided in text). |
| Software Dependencies | No | For all the compared MARL algorithms, we use their default suggested hyperparameter settings in EPy MARL2 (i.e., VDN [39], QMIX [31], MAPPO [46]) or their official implementation (i.e., HASAC [22]). |
| Experiment Setup | Yes | Table 3: The hyperparameters of the compared algorithms in the MOEA/D environment, and "-" means the certain algorithm does not have that hyperparameter. Hyperparameter SADN ACE SAQL VDN QMIX MAPPO HAPPO HASAC Hidden layer size 64 64 64 64 64 64 64 64 Learning rate 1e-4 1e-4 1e-4 1e-4 1e-4 3e-4 3e-4 5e-4 Batch size 32 32 32 32 32 10 10 10 Discount 0.99 0.99 0.99 0.99 0.99 0.99 0.99 0.99 Target update interval 200 200 200 200 200 200 200 50 Number of steps to look ahead 1 5 5 20 Entropy coef 0.01 0.01 - Grad norm clip 10 10 10 10 10 10 10 - |