- [[Learning the identity map]]
- It was found that one-hot vectors of arbitrary dimension can be embedded in 2-dimensional space in a form that is identifiable by the perceptron
  • Discrete states can be represented by one-hot vectors

  • This means,

  • The state is considered a two-dimensional vector.

  • I had a related thought a long time ago: Vectorization of states.

    • At this time I was thinking of constructing a direct

    • Why did you want to state vectorize?

    • I wanted the state transitions to be acquired and improved through learning.

    • Learning State Transition Diagrams

    • This approach is to create training data from a given state transition diagram and train it with a multilayer perceptron.

    • I want to replace existing state transition diagrams written in hard code with neural nets.

    • This is the same concept as in [Replace conditional clauses in if statements with 2-class classification

  • State usually [recurrent

    • If this recurrent part of the system does not need to be human observable, there is no need to make it discrete.

    • Then you can turn it in a compressed vector state.

    • image

    • Reinforcement learning for continuous state may also be relevant


This page is auto-translated from /nishio/状態のベクトル空間への埋め込み 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.