Vectorization of states

• A condition is often thought of as an element (symbol) of a set of discrete. Consider extending this to vector space.

In that case the transition probability is a function of the vector to vector probability distribution

• very
• Represented by several representative points in k-means fashion

Note on synthesis method

• superscription
  • $v\leftarrow v_{new}$
• conversion
  • $v\leftarrow A{v}_{now}$
• synthetic
  • $v\leftarrow v_{new}+A{v}_{now}$
  • Override when A=0
  • When vnew=0, conversion

I want the matrices and state vectors to be acquired by learning.

• Press the “Weird” button when you get a weird reaction.
  • Move to the second nearest representative point
  • Make negative examples training data.
  • (2019 addendum)
    • The idea is to keep representative points even after the state is embedded in the vector space
    • A representative point is just “a set of things that humans were able to verbalize in the first stage, saying, ‘This is the kind of state of affairs that exists.
    • When discretizing after vectorization, one could Voronoi division at the first representative point, or one could k-means on the new distribution.
• Positive response
  • Is direct action the only positive example?
  • Elicit that action as soon as possible.
    • reinforcement learning

---

This page is auto-translated from /nishio/状態のベクトル化 using DeepL. If you looks something interesting but the auto-translated English is not good enough to understand it, feel free to let me know at @nishio_en. I’m very happy to spread my thought to non-Japanese readers.