Python ACTR: A New Implementation and A New Syntax

1
Python ACT-RA New Implementation and A New
Syntax

• Terrence C. Stewart and Robert L. West
• Carleton Cognitive Modelling Lab
• Institute of Cognitive Science, Carleton
  University, Canada
• ACT-R Workshop 2005

2
Vision of the Future of ACT-R

• A standard tool for modelling cognition
• As widespread and widely used as basic statistics
• Easy for any researcher to put together a model
  to test their theory about a particular aspect of
  cognition
• Easy to develop new ACT-R modules
• Easy to collect data from ACT-R models, do
  multiple runs, interact with various user
  interfaces, other ACT-R models, and other
  cognitive models.

3
How can we achieve this?

• ACT-R 6
• Alternate implementations
• jACT-R
• Very few cognitive scientists are learning Lisp
• Not knowing Lisp makes it difficult to create the
  experimental framework around an ACT-R model
• Having ACT-R available in other languages means
  more people will be tempted to make use of it
• Alternate syntax
• Current syntax is based in Common Lisp
• More familiar syntax may entice more users

4
System Design

• Python
• Cleanly supports all of Lisp's essential
  features except macros lthttp//www.norvig.com/pyt
  hon-lisp.htmlgt
• It's Lisp with one fixed syntax that has been
  optimized for fast development, clarity, and ease
  of learning (executable pseudo-code)
• One implementation, extensive standard libraries
• Virtual Machine language, similar speed as Java

5
def incrementSum( goal'add ?count ?sum',
retrieval'count-order ?sum ?newsum')
goal(sumnewsum) retrieval('count-order ?count
?')
(P increment-sum goalgt
isa add count
count sum sum
retrievalgt isa count-order first
sum second newsum gt goalgt
sum newsum retrievalgt isa
count-order first count )
6
import actr class Count goalactr.Buffer
specify the modules
retrieveactr.BasicMemory productionactr.BasicP
roduction memory"""count 0 1, count 1 2,
initialize count 2 3, count 3
4, declarative memory
count 4 5, count 5 6, count 6 7,
count 7 8, count 8 9, count 9 10"""
def start(goal'count-from ?start ?end
starting') define the retrieve('count
?start ?next') productions
goal('count-from ?start ?end counting')
def increment(goal'count-from ?x !?x counting',
retrieve'count ?x ?next')
print x retrieve('count ?next ?nextNext')
goal(xnext) def stop(goal'count-from ?x ?x
counting') print x goal('count-from ?x
?x stop') modelactr.ACTR(Count)
create the model model.goal('count-from 2
4 starting') set the goal
chunk model.run()
run the model
7
Chunks

• Lisp ACT-R chunks
• (chunk-type count-order first second)(a isa
  count-order first 2 second 3)
• Python ACT-R chunks
• count 2 3
• No slot names (use order instead)
• No special 'isa' slot (usually first slot)
• No chunk-types
• Why the big change?
• To reduce the idea of the semantic value of slots
• To reduce the value value confusion between
  slot names and variables

8
Existing Modules

• Declarative Memory
• BasicMemory no learning (ESC F)
• FastMemory optimized (OL T)
• FullMemory non-optimized (OL F)
• Productions
• BasicProduction (PL F)
• PGCProduction (PL T)
• CompilingProduction with compilation (EPL T)
• Others
• Buffer a simple buffer with a do-nothing module
• Used for the Goal Buffer
• Total size 20.9kB, 750 lines

(No partial matching or spreading activation yet)
9
Future Work Open Design Questions

• Production compilation
• We currently just do the RHS of the two base
  productions, so memory requests still happen,
  even though their response is not used
• Can we simplify the constraints on when
  productions can be combined? Do we need the
  limits imposed by Lisp ACT-R?
• The Goal Buffer
• The goal buffer is not special in Python ACT-R
• Can have productions which don't match on the
  goal buffer
• Is this a way of doing reactive models?

10
Future Work Open Design Questions

• Slots with no names
• Any aspects of ACT-R which require named slots?
• Does it cause development confusion?
• Does it get rid of people making chunks with many
  slots?
• Does not requiring an 'isa' match change things?
• Temporary slot names
• Currently, if a variable is bound to a particular
  slot, that variable also becomes a temporary name
  for that slot within that production
• Meant to simplify a RHS rule that just changes a
  few slot values

11
Future Work More Modules

• Partial Matching
• With an easier way of specifying matching values
• Spreading Activation
• How do we work with slots referring to other
  chunks?
• This is possible now, but needs a clear syntax
• Production learning via other Reinforcement
  Learning systems
• Instead of PG-C, use Q-Learning or other modern
  reinforcement learning systems (Gray et al, 2005)
• ACT-R/PM
• Integrated with the Tcl/Tk and wxPython gui
  libraries
• SOS for Python ACT-R
• A simple interface to manipulate objects

12
Discussion Questions

• Is this useful to any of you?
  lthttp//ccmlab.ca/actrgt
• What would it need in order to be more useful?
• What divergences from Lisp ACT-R seem the most
  dangerous? The most interesting?
• Does it fit with ACT-R 6?
• Is having two different syntaxes for ACT-R more
  confusing or less confusing?
• Will this help people researching ACT-R?
• Will this help people who just want to use ACT-R
  to create models?