added toysgd benchmark - Githubissues

automl / DACBench

A benchmark library for Dynamic Algorithm Configuration.

https://automl.github.io/DACBench/

Apache License 2.0

28 stars 13 forks source link

added toysgd benchmark #103

Closed benjamc closed 3 years ago

benjamc commented 3 years ago

Please have a look. :-) Especially on the bounds for the coefficients and the initial x (not sure about that).

TheEimer commented 3 years ago

Initial x as 0 makes sense, I think. Without haing every run it, I can't really say if the coefficient range is good or not, but it's super easy to adapt, so I think that's good. So thanks for the benchmarks, I'll merge as soon as you adressed my comment (second one is optional).

TheEimer commented 3 years ago

Any updates here, @benjamc ? I found another thing to fix in the meantime, the reset doesn't seem to return a state. I missed that, but it's definitely not desired behavior ;D

TheEimer commented 3 years ago

Also, we may want to set an upper limit for values. The lr get huge pretty fast with actions > 1

TheEimer commented 3 years ago

Changes from my side:

limited actions to stay below 1
made reset return a state
initialized lr with 0 (might not be smart at all, but None is worse)

Things where input would be nice:

how can we initialize the lr without giving the agent a nan in the reset state?
is 1 a good limit for the actions?

As soon as we settled those two, we can merge

benjamc commented 3 years ago

[lr] setting the learning rate to 0 as a starter should be fine. this way the agent is definitely forced to do something :D and the agent does not set the learning rate itself but the log learning rate . I am not sure if this is really good or if we should set the learning rate to a really small value because the agent can never achieve a learning rate of 0 because of log.
[actions] do not limit actions