Game Playing Agents
Reinforcement-learning agents trained on a laptop CPU, and one traffic signal controller, all scored under the same protocol: a named baseline, disjoint selection and reporting episodes, and a gain that only counts if it clears two standard errors of the difference.
Game Agents
Eight agents on eight real downloadable games — DOOM, Crafter, Brogue CE, a MuJoCo robot, CarRacing and three Atari titles. Eleven measurements, four of which do not clear their bar and are published anyway.
Traffic Agents
Nine signalised junctions, five controllers, one shared PPO policy. Saves delay against the best hand-written controller at two of three demand levels, and is inside its own noise at the third.
How every number here was measured
- The checkpoint ladder is scored on one set of episodes; the best checkpoint is re-scored on a different set.
- A gain counts only if it exceeds two standard errors of the difference.
- Baselines are named. A random policy is scored as a floor and labelled as one, and is never allowed to win.
- Results that did not clear their bar stay in the tables, with the same error bars. That is what makes the rest of them worth reading.