Multi-agent Dynamic Algorithm Configuration

• DOI: 10.48550/arxiv.2210.06835
• Publication Date: 2022-01-01
• Authors: Ke Xue; Jiacheng Xu; Lei Yuan; Miqing Li; Chao Qian; Zongzhang Zhang; Yang Yu
• Keywords: Hyperparameter; Benchmark (surveying)

Automated algorithm configuration relieves users from tedious, trial-and-error tuning tasks. A popular paradigm is dynamic (DAC), in which an agent learns policies across instances by reinforcement learning (RL). However, many complex algorithms, there may exist different types of hyperparameters, and such heterogeneity bring difficulties for classic DAC uses a single-agent RL policy. In this paper, we aim to address issue propose multi-agent (MA-DAC), with one working type hyperparameter. MA-DAC formulates the multiple hyperparameters as contextual Markov decision process solves it cooperative (MARL) algorithm. To instantiate, apply well-known optimization multi-objective problems. Experimental results show effectiveness not only achieving superior performance compared other approaches based on heuristic rules, multi-armed bandits, RL, but also being capable generalizing problem classes. Furthermore, release environments paper benchmark testing MARL hope facilitating application MARL.