MDGPS - debug step rule

cbfinn / gps

Guided Policy Search

http://rll.berkeley.edu/gps/

Other

594 stars 239 forks source link

MDGPS - debug step rule #37

Closed wmontgomery4 closed 8 years ago

wmontgomery4 commented 8 years ago

This PR fixes a bug in the step size adjustment for MDGPS, along with some hyperparameter tweaks to improve 'mjc_mdgps_example' variants.

wmontgomery4 commented 8 years ago

I'm running this now on mjc_badmm_example to test the random_seed/quit_on_end stuff. Assuming it works, we'll be ready to merge.

wmontgomery4 commented 8 years ago

Okay, this is ready to merge. Contains:

Bug fixes for stepadjust in 'MDGPS
Hyperparameter tweaks for mjc_mdgps_example
Addition of 'random_seed' hyperparam (goes in config), and --quit command line flag

I chose to add the --quit option as a command line flag rather than a hyperparam since it will mostly be used that way inside scripts and it would be better to not have to include it in the hyperparams.py file. This adds a second argument to the GPSMain object and feels a little hacky, but I think it's the better choice for now.

cbfinn commented 8 years ago

LGTM

cbfinn commented 8 years ago

By the way, one way to perhaps make the --quit option a bit less hacky, would be to set

hyperparams.config['quit_on_exit'] = args.quit

Then, you only have to pass in hyperparams.config, and can read in that option.

Also, this won't quit in the event of an error. I think the only way to quit in the event of an error would be to put error-prone code (or just all of the code) in a try/catch statement.