Significant deep reinforcement learning advances, 2013–2026. Core rows introduced a new mechanism or landmark result; refinements port, combine, or fix an existing one. An algorithm lists its primary domain first and, in smaller type, the others it was demonstrated on; the domain filter matches any of them. The page opens on the Atari scores of every discrete-action algorithm; switch to the table for the full atlas, and click a year, family, or column header to filter and sort.
| Year▲ | Algorithm | Family | Domain | Core breakthrough | Atari |
|---|
Four panels: the standard 200M-frame budget, the runs that went far past it (1B to 90B frames, so MuZero at 20B sits above MuZero Reanalyse at 200M without being the better method at equal data), the 100k benchmark and a 40M-frame 20-game run are not on the same axis. Every number was checked against the paper table its source line names (2026-09-13); grey bars are the reference runs a paper measured alongside its own method, so a panel can be read within one protocol. Human starts, sticky actions and game counts differ between papers, so read differences under a factor of ~1.3 as noise: Rainbow is 223% in its own table, 153% with human starts and 147–150% under Dopamine's sticky actions. Median or mean is the toggle above; hover a bar for the other statistic and the source. Game-specific results (Montezuma's Revenge, Procgen) sit in the table's Atari column instead.
Years are first public release (arXiv or blog), not conference year. LLM post-training methods (RLHF, DPO, GRPO and descendants) are deliberately excluded. 2026 entries come from the RLC 2026 pre-proceedings (RLJ vol. 7); their tiers reflect the novelty of the mechanism, not yet proven influence.