Why Greatness Cannot be Planned

tags: [ src:book , artificial_intelligence ]

source: reddit

• book is overkill (haven’t read it), but essentially:
• examples in reinforcement learning where if all you do is optimize for an explicit objective, then the algorithm will oftentimes do silly things
• instead of moving towards the objective, why not bounce around and looking for novelty, so you get serendipity (and creativity)11 which I assume how he is able to generalize this principle in CS to the real world..
  • feels like the whole bandit setting exploration vs exploitation problem
• this is usually in the context of evolutionary algorithms, which usually optimize for some fitness criteria (at each epoch).

on a more positive note:

• this relates to what I’ve been thinking about recently, which is, how do we move forward from the neural network training paradigm.
  • does it make sense to have an objective function, and training data?