Experiments and Logging

My high-level goal over the last week or so involved finishing the write-up for trajectory collection for constrained and unconstrained agents.

Immediately upon starting training, I butted up against the question of how and when to log experiments and agent performance. Indeed, it is important to be able to determine which, among the agents one has trained, is considered an “expert.” This involves a dual process:

  1. saving good hyperparameter sets for future training exercises
  2. hand-picking agents from observed performance in the set environment

Out of the box, OpenAI Gym sets the lambda user up with great reporting at the epoch level, but as a visual learner who anticipated lots of future experiments, this would be lacking. I decided to adopt Weights and Biases (W&B) into my workflow for two reasons:

  1. track experimental configurations and log performance - wandb.config() is the only command you need to log hyperparameter settings and it is compatible with ArgParser - W&B reports immediately give you the ability to mix data with commentary on results and transitions towards newer experimental sets was a top selling point for my project
  2. the ease with which results can be shared and evaluated

An unexpected boon from this choice was the Hyperparameter sweep functionality, that can be set-up with uncanny ease. Indeed, all you need to do is select the sweep mode: [“grid”, “bayes”] and set your parameter values in the acceptable formats.

The output you receive is displayed on the Weights and Biases page and can be downloaded in report form.

NameRuntimeHostnameNotesStateSweepTagscost_gammacost_limgammahidlseedstepssteps_per_epochcost ratecumulative costepoch
firstsweep118tyna-server-finishedyy302nxi0.99100.9864245640000.099788461538462518912
firstsweep427tyna-server-finishedyy302nxi0.99100.9864212340000.0737751475549
firstsweep411tyna-server-finishedyy302nxi0.99100.986429940000.065651313049
firstsweep375tyna-server-finishedyy302nxi0.99100.9864199940000.077651553049
firstsweep336tyna-server-finishedyy302nxi0.99100.9864145640000.0701851403749
firstsweep361tyna-server-finishedyy302nxi0.99100.986419940000.0637051274149
rose-sweep-10-crashedyy302nxi0.986410
secondrun360tyna-server-finished100.99646420400040000.0756851513749
pretty-bird-1330tyna-server-failed00.07347549
Tyna Eloundou
Tyna Eloundou
Scholar

My research interests include alignment (inner and outer variants), safety and AI creativity.