Mirror of forum.nim-lang.org

5120 :: Marvin Minsky frame model /extended with ordered storage/

• dponyatov (orginal) [2019-08-23T13:43:13+02:00] view original

  • https://en.wikipedia.org/wiki/Frame_(artificial_intelligence)
  • https://web.media.mit.edu/~minsky/papers/Frames/frames.html

  What is the right code for implementing the class hierarchy of frames?

  I look at the base Frame class as a universal data (and code) type, can hold both single scalar values, and nested elements, all of them can be attributed by arbitrary slots. Making oriented graphs of such nodes with types inherited from the base Frame, we can model any type of data. Some subset of frames can be treated as executable so we can implement a homoiconic computing system.

  type
      Frame = ref object of RootObj
          typ: string
          val: string
          slot: Table[string,ptr Frame]
          nest: seq[ptr Frame]
  
  let nst = Frame(val:"nested")
  let hello = Frame(typ:"hello",val:"world",nest:@[nst])
  
  echo hello[]
  Run

---

• mratsim (orginal) [2019-08-23T20:35:51+02:00] view original

  IMO, those links are too theoretical and you most probably need domain knowledge and actual use cases to know the best data structures to represent Frames.

  Looking around you can probably re-use OO techniques though: https://en.wikipedia.org/wiki/Frame_language#Comparison_of_frames_and_objects.

  Couple of notes:

  • Your proposal will be bottlenecked by string processing, at least typ should be an enum
  • ptr Frame should be just Frame as Frame is a ref object.
  • I don't know the associated procs/use-cases but all those heap allocations will lead to memory fragmentation, bad memory locality and poor performance if you need to create/destroy Frames in tight loop.
  • That Frames theory seems to heavily suffer from the curse of dimensionality

  My advice to go forward:

  • define what a minimal proof-of-concept should be able to do
  • implement it
  • refine the base data structure so that building that use-case from that data structure is easy.
  • add new use-cases, rinse-and-repeat.

  Check performance from time to time on non-toy inputs.

---