2022/08/31 by Quentin Gallouédec, Gallouédec, Quentin, Emmanuel Dellandréa +1 · 3 citations
Biochemistry, Genetics and Molecular Biology · #Cell Image Analysis Techniques
paper · pdf · doi:10.48550/arxiv.2208.14928
In this paper, we introduce Latent Go-Explore (LGE), a simple and general\napproach based on the Go-Explore paradigm for exploration in reinforcement\nlearning (RL). Go-Explore was initially introduced with a strong domain\nknowledge constraint for partitioning the state space into cells. However, in\nmost real-world scenarios, drawing domain knowledge from raw observations is\ncomplex and tedious. If the cell partitioning is not informative enough,\nGo-Explore can completely fail to explore the environment. We argue that the\nGo-Explore approach can be generalized to any environment without domain\nknowledge and without cells by exploiting a learned latent representation.\nThus, we show that LGE can be flexibly combined with any strategy for learning\na latent representation. Our results indicate that LGE, although simpler than\nGo-Explore, is more robust and outperforms state-of-the-art algorithms in terms\nof pure exploration on multiple hard-exploration environments including\nMontezuma's Revenge. The LGE implementation is available as open-source at\nhttps://github.com/qgallouedec/lge.\n