Production Rule Learning

The Basic Idea

• production compilation
  • discuss how [~ new production rules] are learned
  • 2 production rules that apply in succession —> a single rule
• In the pair task: (for the pair zinc & 9 )
  • PRODUCTION 1 : see =val, retrival (abstract pattern matching )
  • PRODUCTION 2 : retreived =ans, press =ans (abstract pattern matching )
  • COMBINED PRODUCTION : see zinc, press 9 (specific)

Forming a new production

• 2 sets of conditions —> a single set of tests
• 2 sets of actions —> a single set of actions that has the same overall effect
• Motor Style Buffers
  • manual buffer & visual buffer
    • they never hold a chunk
    • only receive request
    • only tested through query
  • Condition of compilation: - 1st : request bufferA , 2 nd : request bufferB ✅ - 1st : request bufferA , 2 nd : request bufferA ❌ - 1st : request bufferA , 2 nd : query bufferA state busy　✅ - 1st : request bufferA , 2 nd : query bufferA state free (other than busy) 　❌
• Perceptual Style Buffers
  • visual-location, visual, aural-location, aural
    • they will hold chunks generated by their modules
  • Condition of compilation:
    • 1st : request bufferA , 2 nd : query bufferA state free (other than busy) ❌
    • 1st : request bufferA , 2 nd : request bufferA ❌
    • 1st : request bufferA , 2 nd : test contents of bufferA ❌
• Retrieval Style Buffers
  • retrieval
    • more predictable and offers a chance for economy
  • Condition of compilation:
    • 1st : request retrieval , 2 nd : test the results of the retrieval ✅
      • specialize the composed production
    • 1st : request retrieval , 2 nd : query for a retrieval failure ❌
• ** Goal and Imaginal Style Buffers**
  • goal, imaginal
  • first prodcution does not make a request
    • before compilation : C1 –> A1 , C2 –> A2
    • after compilation : C1 + (C2 - A1 ) —> A2 + (A1~A2)
  • first production makes a request
    • The second production does not also make a request
      • 2nd buffer test can be deleted
      • C1 –> A1 (modifycation) & A2+(N1~A2) (new request)
    • The second production also makes a request
      • can not be composed ❌
  • Difference between goal and imaginal - the use of query
    • imaginal can be busy
    • goal never be busy
      • For goal:
    • never query goal buffer, otherwise the compilation will be blocked
      • For imaginal:
    • if the 1st has a request then the 2nd must
      • test the buffer for state busy
      • not also make a request
      • Detailed Info : compilation.xls

Utility of newly created productions

• New = Old1 + Old2
• whenever New could apply, Old1 can also apply (Old1 is more 汎用的)
• Utility of New : :nu default = 0
• Each time New is created, its utility is updated with a rewad
  • reward = Utility of Old1
• New lose and Old1 fire —-> New and Old1 is very close —> occasionally New is selected –> New surpass Old1 (because it is faster and received more reward)
• The speed of learning is controlled by the parameter :alpha

  • If it is set to 1, productions will typically get very good values immediately and likely be tried on the first opportunity.

Learning from Instruction

• usually, instructions of an experiment is learned by the subject in the warming phase.
• in ACT-R, they are the productions written in the code
• operator : pre, action, arg1, arg2 , post
• Six Instructions of the Pair Experiment ( Tutorial/Unit7/paired-learning.lisp)
  1. At the start read the word and create an encoding of it as the stimulus
  2. After reading the stimulus try to retrieve an associate to the stimulus
  3. Test whether an item has been recalled and if it has not then just wait
  4. If an item has been recalled type it and then wait
  5. Store the response you read with the stimulus
  6. This trial is complete so start the next one
• The model has only 3 production : retrieve-operator , read-arg1 and encode-arg1
  • neither of them were specific productions of the pair task

  (p retrieve-operator
   	=goal>
   		isa task
   		state =state
   		step ready
   ==>
   	+retrieval>
   		isa operator
   		pre =state
   	=goal>
   		step retrieving-operator    ;
   )

• with production compilation on:
  • The paraemter is :epl
• with PC on, increases the speed-up over trials in the recall time.
  • Correlation 0.974 –> 0.992
• To see new productions in the trace:

  • ( :ptc t)