bonsai train

New Contributor


What parameters of the graph can be changed so that when training is interrupted, it does not start again?

2 Replies
By "parameters of the graph", do you mean "what aspects of the inkling document can be modified"? In general, it is any aspect that affects reproducibility of training. For example, adding a comment, adding a new concept, adding a new lesson, renaming a variable, constant or function, etc. does not affect the reproducibility of what has already been trained. In contrast, modifying an objective within a goal, adding a new objective, modifying a reward function, modifying state or action types, etc. do affect reproducibility and therefore reset training.

When you make a change to the graph (either in the inkling text editor or the visual editor), you can tell immediately whether the change will result in retraining because the web client will tell you as much. You can easily choose to undo the change if you want to avoid the retraining. That's the easies way to tell whether a particular change will cause retraining of some portion of the graph.

Also keep in mind that any time retraining is needed, the Bonsai client creates a new version, preserving the training in the older version. So you can always go back to that older version if you want to continue experimenting from that point.

@erictr Yes, you understood me correctly, although I may have expressed myself poorly. Thanks for the prompt reply, it was helpful!