atune live demo: tuning a real learner
By default, atune tunes a real reinforcement learner in your browser: each dot is a full training run of Augmented Random Search on CartPole-v1, and the strip below replays the best policy so far. With a scheduler, atune stops or forks runs from their learning curves. Pick another problem, any of nine samplers, a scheduler, the number of trials and the seed; the same settings repeat the same run. The model map is a Gaussian process fit to the trials.
Problem and optimizer
Run
View
Write f with x, y, numbers, + − * / ^ and parentheses, pi and e, and sin, cos, tan, asin, acos, atan, sinh, cosh, tanh, exp, log (natural; also ln), log10, log2, sqrt, abs, floor, ceil, sign, min(a, b), max(a, b), atan2(y, x) and hypot(a, b); 2x means 2*x. Apply (or Ctrl+Enter) restarts the run on it. With drawing on, a click or drag on the heatmap appends Gaussian valleys or hills to the formula, each 0.4 of the function's current range deep.
- completed trial (older ones fainter)
- stopped or paused early
- lowest completed value
- newest trial
- global minimum (custom: lowest grid point)
- model's lowest prediction or highest expected improvement
- TPE's good set
- CMA-ES search distribution
Problem and optimizer
Trials and model
Trials, newest first
| Trial | x: log10 step size | y: log10 noise | Value | Best so far | Outcome | Trial seed |
|---|