3 ms·
I can definitely understand wanting to classify Pavlov as supervised learning. I think it's a murky issue because supervised and reinforcement are very closely
by nepstein 11y ago
I can definitely understand wanting to classify Pavlov as supervised learning. I think it's a murky issue because supervised and reinforcement are very closely related (and it is frequently possible to reframe problems of one type as problems of the other depending on what type of model one uses).
My two main reasons for going with this name are as follows (if people see issues in my logic, I'm happy to be convinced):
1) Reinforcement learning gets its name from the behaviorist psychology concept of reinforcement in which an agent's actions are met with rewards in order to shape that agent's future behavior. This is precisely the kind of response conditioning that Pavlov is well known for.
2) The key difference is what the training data look like.
In a supervised learning problem, the training data are input/output pairs (a stimulus and an appropriate action).
In reinforcement learning, the training data are input and reward pairs (an action and the reward applied to that action).
I would argue that Pavlov's experiments more like the latter case - the dogs are not shown 'this is the correct action for this stimulus', they are shown 'this is the reward for this stimulus'.
- graycat 11y agoMarkov decision processes is an old, mature, at times deep, and polished field. Names include R. Bellman, E. Dynkin, R. Rockafellar, D. Bertsekas. There are connections with scenario aggregation, potentials, linear-quadratic-Gaussian certainty equivalence, currents of sigma algebras, the strong Markov property, stopping times, and much more. Can we be more clear on just what the Markov processes involved actually are and, then, how they are to be used?
- nepstein 11y agoThe README links to a blog post (http://nepste.in/jekyll/update/2015/02/22/MDP.html http://nepste.in/jekyll/update/2015/02/22/MDP.html) which details how the library is implemented from the definition of a MDP. For a more rigorous treatment, Andrew Ng's notes (http://cs229.stanford.edu/notes/cs229-notes12.pdf http://cs229.stanford.edu/notes/cs229-notes12.pdf) are an excellent resource.
- graycat 11y agoYour first reference is good enough -- it's okay. All I saw was the Github page of gibberish -- I don't use Github whatever the heck it is. But your URL was fine. So, the work is a relatively routine application of classic work from optimization going way back, e.g., to Bellman. The "Reinforcement learning" terminology looks like a new label for some quite ancient wine. I've wondered what machine learning had that was good and new, and so far I've seen some that is good but not new and some that is new but not good. For an application, it would be good to justify the Markov assumption, that is, that the past and future of the process are conditionally independent given the present. For a more detailed treatment, I'd recommend, say, E. B. Dynkin and A. A. Yushkevich, 'Controlled Markov Processes'.
- defen 11y agoHi - in a previous comment you mention a paper you wrote that describes a distribution-free multivariate anomaly detector (this is the comment: https://news.ycombinator.com/item?id=9580929 https://news.ycombinator.com/item?id=9580929) Would you mind emailing me a copy of it please? Address in profile. Thanks in advance!