Mastering Atari 2600 Games with Discovered Options
Wayfarer is presented as an Atari agent that discovers its own reusable options from pixels and uses them to learn hard games faster.
The paper says the method learns temporal abstractions online with Laplacian representation learning, without relying on handcrafted or symbolic state descriptions. Those options are used for control and are reported to improve exploration, credit assignment, and generalisation to unseen settings. The authors claim state-of-the-art results among single-stream agents on the hardest Atari 2600 games, with especially large gains on Montezuma's Revenge and Private Eye. Source: ArXiv · AI/CL/LG's note
The paper says the method learns temporal abstractions online with Laplacian representation learning, without relying on handcrafted or symbolic state descriptions. Those options are used for control and are reported to improve exploration, credit assignment, and generalisation to unseen settings. The authors claim state-of-the-art results among single-stream agents on the hardest Atari 2600 games, with especially large gains on Montezuma's Revenge and Private Eye. Source: ArXiv · AI/CL/LG's note
score 5