Activation of Chunks and Base-Level Learning

• 
• Introduce the subsymbolic part of ACT-R

1. Introduction

---

• enable subsymbolic components:

  (spg :esc t)

2.Activation

---

• any chunk associated with a value : activation
• reflects :
  • the degree to which past experiences
  • the degree to which current context indicate that chunk will be useful
• retrieval threshold:

• (sgp :rt -0.5)

• sets the minimum activation a chunk can have and still be retrieved.

• Activation = base level + context + noise

3.Base-level Learning

---

• 
• d : decay rate
• n : The number of presentations for chunk i.
• t : The time since the jth presentation.
• two types of presentation:
  • initial entering into DM

    add-dm

  • When the chunk is cleared from a buffer. ( will be add into DM)
    • merge with another that is already in DM
    • If a chunk has the same set of slots and values as a chunk which already exists in the DM
    • preexisting chunk in DM is credited with a presentation.
    • repeatedly attending the same visual stimuli would result in strengthening a single chunk that represents that object. (重复看到同一个刺激会加强对于该刺激的记忆)

4.Optimized Learning

---

• take time for computation
• an approximation approach
  • turning on the optimized learning parameter - :ol.　　(default is on)
  • 使用条件： the presentations are approximately uniformly distributed over the time since the item was created.
• 
• n: The number of presentations of chunk i.
• L: The lifetime of chunk i (the time since its creation).
• d: The decay parameter.

5.Noise

---

• logistic distribution characterized by a parameter s
• 
• two source of noise
  • permanent noise (s is set by :pns))
  • instantaneous noise ( recomputed at each retrieval attempt ) ( s is set by :ans)
  • typically turn permanent noise off

    (sgp :pns nil) (sgp :ans t)

6.Probability of Recall

---

• 
• In fact, when τ = Ai, the probability of recall is .5
• s controll the sensitivity of recall
  • s 接近0 ： 概率在0，1之间反复横条
  • s 较大 ： 平滑的sigmoid曲线

7.Retrieval Latency

---

• The activation of a chunk also determines how quickly it can be retrieved.
• 
• A: The activation of the chunk which is retrieved.
• F: The latency factor (set using the :lf )).
• if no chunk match or no chunk exceed threshold, then time it takes for the failure to be signaled is :
  • 
  • τ: The retrieval threshold.

8. The Paired-Associate Example

---

• paired stimuli prented for 5s each
  • word —– number
  • zinc —- 9
  • xray —- 8
  • girl —- 7
  • boy —- 6
• first time presented: learn the pairs
• 随着练习次数增加（trial），回答准确率和反应速度上升
• モデルの流れ図：
• 
• detect-study-item 和 associate production: 加强记忆

• (p detect-study-item
    =goal>
        isa goal
        state read-study-item
    =visual-location>
    ?visual>
    	state    free
   ==>
   	+visual>
    	cmd   move-attention
    	screen-pos =visual-location
    =goal>
    	state    attending-target
  )
  

• (p associate
    =goal>
      isa      goal
      state    attending-target
    =visual>
      isa      visual-object
      value    =val
   =imaginal>
      isa      pair
      probe    =probe
    ?visual>
      state    free
  ==>
   =imaginal>
      answer   =val    ; 将 probe 和 answer 关联起来
   -imaginal>          ; 关联好的pair存储到DM中
   =goal>
      state    start  
   +visual>
      cmd      clear
  )
  

• The associate production fires after the model reads the number which is associated with the probe
• 如果没有开启 activation component，则从第二个trial开始，模型的正确率将始终保持100%，且反应时间也没有变化

• 

• 开启 activation component 之后，もっと人間っぼく振舞う
• 

9. Parameter estimation

---

• 4 parameters should be set
  • the retrival threshold = -2 (default)
    • 
  • the instantaneous noise = 0.5 (determines how quickly probability of retrieval changes as we move past the threshold.)
    • 
  • latency factor is set at 0.4. (determines the magnitude of the activation effects on latency.)
    • 
  • decay rate for base-level learning is set to the value 0.5 (recommend it be set for most tasks )
    • 

10. The Activation trace

---

(sgp :act t)

11. The :ncnar Parameter

---

• normalize chunk names after run
• do not affect model’s performance
• affect actual time it takes to run the simulation, turn off can save time