@InProceedings{RehlingHofstadter1997, author = {John Rehling, Douglas Hostadter} title = {The parallel Terraced Scan: An Optimization for an Agent-Oriented Architecture} proceedings = {IEEE International Conference on Intelligent Processing Systems} year = {1997} pages = {900--904} }
Definition:
According to the authors, the parallel and random nature of the processing and search mechanisms are similar to the way human cognitions work.
Parallel terraced scan is designed to explore many possibilities at a time, but to devote more computation to promising directions in search.
The agent-oriented architecture that was optimized was composed of
4 parts [pg.1]:
-A workspace where the letter is parsed
-A Conceptual Memory, which is a network containing nodes for each
possible role for the parts of letters and a role set, BR>-Program
attempts to label the parts
-Wholes are found through spreading activation
-When the temperature is low enough the program halts
The authors had two goals in optimizing the architecture in order to implement parallel terraced scan: 1) increase speed, and 2) increase accuracy.
There were six optimizations that took place [pg.3]:
-Re-parsing of letters clears all activation from previous letter
parse in conceptual memory so that leftover activation does not
interfere with current processing
-Introduction of gestalt codelets which run when letter has been
parsed and activates only those wholes which are legitimate
possibilities
Simulators contain a lot of multimodal information.
-Elimination of a codelet called R-Role checker which allowed for
poor matching wholes to remain highly activated for arbitrarily long
periods of time and this
reduces the number of wholes that need to be processed
-When activation spreads from a node in conceptual memory, the
node is only allowed to retain a small fraction of the previous
activation
-The phases of loosening (decreasing stringency of requirements)
was made to be probabilistic
-Parts can be assigned roles through the influence of roles that
already have a high activation level within memory
To test the performance of the system a set of 544 gridforms was
used and a set of 52 was used for debugging [pg.4].
The set was split into 2 groups: EASY (388 forms) and HARD (156
forms). The performance measure was the amount of letters recognized.
To test the improvements that the parallel terrace scan caused the
amount of codelets ran was measured.
Results:
Set; Codelets/Run Correct
%
Old New Old New
TEST 1381.3
624.8 82.6 85.3
PLAIN 770.2 390.9 95.7 96.8
HARD
2901.1 430.2 49.9 56.8
As can be seen from the data introduction of a parallel terraced
scan allows for significant increases in both speed and performance.
The main reason for the increased performance of a system using a
Parallel Terraced Scan is that it is able to take advantage of prior
computations.
The way that the program is able to do this is to use the Gestalt
Codelete, which is a small initial investment in computation can have
large benefits in total run time.
The use of a system such as this is encouraging because it shows a increase in computational efficiency and it uses a mechanism that could have a have a role in human cognition.