Only that “Exploration vs exploitation trade-off” is not an algorithm. Reinforcement learning (RL) is pretty much “non-algorithmic” (as Pei Wang would say). ETA: there are specific algorithms in RL (and in—related—planning and game playing), but the “trade-off” is a concept; it sure needs to be expressed algorithmically but is it fair to give credit to “algorithmicality” in this case?
Right; when I say “I’m not systematic about it” I mean that I don’t purposefully follow a specific algorithm. I would probably benefit from being a bit more systematic, but for the moment, I’m merely trying to “train my intuition”.
I would hope that all these algorithms would be applied “non-algorithmically” in Pei Wang’s sense—that is, the ideas from the algorithm should interact dynamically with the rest of my thought process.
Reinforcement learning is pretty much “non-algorithmic”
I’m rather certain I could implement reinforcement learning as an algorithm. In fact, I’m rather certain I have done so already. If I can point to an algorithm and say “look, that’s a damn reinforcement learning algorithm” then I’m not sure how meaningful it can be to call it “non-algorithmic”.
I concede, RL is a prototype example of algorithmic learning problem. The exploration vs exploitation trade-off is something that needs to be addressed by RL algorithms. It is fair then to say that we gain insight into the “trade-off” by recognizing how the algorithms “solve” it.
Only that “Exploration vs exploitation trade-off” is not an algorithm. Reinforcement learning (RL) is pretty much “non-algorithmic” (as Pei Wang would say). ETA: there are specific algorithms in RL (and in—related—planning and game playing), but the “trade-off” is a concept; it sure needs to be expressed algorithmically but is it fair to give credit to “algorithmicality” in this case?
Right; when I say “I’m not systematic about it” I mean that I don’t purposefully follow a specific algorithm. I would probably benefit from being a bit more systematic, but for the moment, I’m merely trying to “train my intuition”.
I would hope that all these algorithms would be applied “non-algorithmically” in Pei Wang’s sense—that is, the ideas from the algorithm should interact dynamically with the rest of my thought process.
I’m rather certain I could implement reinforcement learning as an algorithm. In fact, I’m rather certain I have done so already. If I can point to an algorithm and say “look, that’s a damn reinforcement learning algorithm” then I’m not sure how meaningful it can be to call it “non-algorithmic”.
I concede, RL is a prototype example of algorithmic learning problem. The exploration vs exploitation trade-off is something that needs to be addressed by RL algorithms. It is fair then to say that we gain insight into the “trade-off” by recognizing how the algorithms “solve” it.
It is also fair to say there is an abstract concept of ‘trade off’ that is not itself algorithmic.