Loading ...
Sorry, an error occurred while loading the content.

6372Re: DDFA Source

Expand Messages
  • kenstanley01
    Jun 18, 2014
    • 0 Attachment
      Hi Andy, in response to you points:

      1) Thanks for the link to the paper.  This issue of whether there is a need for some kind of guidance within DDFA reflects the broader debate on whether novelty alone is sufficient within any given problem domain.  There are interesting arguments on both sides of this type of debate and there generally ends up being a tension between two seemingly intolerable pitfalls:  either you allow the objective (i.e. guidance) to pull the search towards deception and bad representation or you completely give up on having any guidance whatsoever and just hope things work out.  However you fall on the spectrum of possible resolutions to this tension, clearly the answer merits careful consideration and is likely domain-dependent.  In that context, DDFA is different form an autoencoder (which is in the end just another algorithm based on minimizing an error) so the particular worries that Snoek et al. address with respect to unguided autoencoders may not apply to DDFA, but there may nevertheless be issues unique to DDFA that lead to a compelling argument about guidance.  Whatever the case is, guidance of any sort (regarding DDFA or anything else) will have to applied delicately and expertly if at all, as is the case with any kind of objective interference in a healthy non-objective divergent process.

      2) I agree that some variations of DDFA that are designed to build new layers would require adding new code or changing code in various NEAT/HyperNEAT/novelty search implementations. That's pretty standard for any new method and like I said we intend to release DDFA code ourselves to make it easier for people to try out variations.  While I completely understand why you would like some more detail on the best alternatives for adding depth, at the moment it's such an open question with so many possibilities that I want to narrow down a bit for myself the best options before speculating about it too much publicly.  For example, the fact that these features are represented by HyperNEAT means there are a lot of interesting possibilities for exploiting geometry as depth goes up.  Once I have a better handle on what's most promising I will certainly be happy to share.


    • Show all 7 messages in this topic