A Quantitative Analysis of Memory Usage for Agent Tasks
Created by W.Langdon from
gp-bibliography.bib Revision:1.7615
- @InCollection{Kim:2008:FER,
-
author = "DaeEun Kim",
-
title = "A Quantitative Analysis of Memory Usage for Agent
Tasks",
-
booktitle = "Frontiers in Evolutionary Robotics",
-
publisher = "IntechOpen",
-
year = "2008",
-
editor = "Hitoshi Iba",
-
chapter = "14",
-
pages = "247--274",
-
address = "Rijeka",
-
keywords = "genetic algorithms, genetic programming",
-
isbn13 = "978-3-902613-19-6",
-
bibsource = "OAI-PMH server at mts.intechopen.com",
-
identifier = "doi:10.5772/5458",
-
language = "en",
-
oai = "oai:intechopen.com:856",
-
relation = "ISBN:978-3-902613-19-6",
-
rights = "https://creativecommons.org/licenses/by-nc-sa/3.0/",
-
URL = "
http://www.intechopen.com/articles/show/title/a_quantitative_analysis_of_memory_usage_for_agent_tasks",
-
URL = "
https://cdn.intechopen.com/pdfs/856.pdf",
-
URL = "https://doi.org/10.5772/5458",
-
DOI = "
doi:10.5772/5458",
-
size = "28 pages",
-
abstract = "The number of states in finite state machines in the
experiments may not be exactly the same as the number
of states that the evolved controllers actually use for
exploration. It specifies a maximum limit over the
number of finite states. Especially for a large number
of states, controllers that do not use all memory
states are sometime evolved even though a given maximum
memory limit is specified. The same can be true of the
genetic programming structure. When the maximum number
of terminal nodes was set up for evolutionary runs,
some best controllers used a smaller number of nodes
than the limit size, or had a redundant expression.
Thus, our analysis of memory states may have a little
discrepancy with the actual usage of memory. Genetic
programming has a high-level representation feature
with a procedural program. When an S-expression is
translated into a finite automaton, it has a main loop
for repeating the action sequence. It often has a
sequential process among internal states until the end
of program is reached. In contrast, the Mealy machine
notation allows transition loops among internal states.
Evolving the FSM controllers can create such loops for
in-between states (from state to state) and more
conditional transition branches. The flexible
representation of the Mealy machine provides more
dynamic property for a given number of states. The
performance difference between the two types of
controllers is due to the characteristics of
representation. To discriminate the performances of a
varying number of internal states, the beta
distribution of success rate or computational effort
was used. We believe that the success rate is a better
criterion for this application, because we are more
interested in the on-off decision of the quality of
controllers with a given evolutionary setting rather
than efficient development of controllers. The
computational effort can be more effective when
strategies to be compared have different computing
costs or when the efficiency is a major criterion in
the evolutionary experiments. An assumption for the
suggested significance test of computational effort is
that each single run has almost the same level of
computing cost. If each run may have a significantly
different computing cost, the estimated computational
effort based on success rate would have a deviation
from the actual effort. The run-time distribution, that
is, the curve of success rate for variable computing
cost provides the characteristics of a given algorithm
and we can easily observe the transition of performance
with run-time. The run-time distribution with its
confidence range would be a useful tool to compare
different algorithms. In the evolutionary computation
research, the performance comparison among evolutionary
algorithms has often used the average performance over
fitness samples or t-statistic. We argue that the
comparison without observing the fitness distribution
may not notice significant difference. The beta
distribution analysis or Wilcoxon rank-sum test would",
-
notes = "Yonsei University, School of Electrical and Electronic
Engineering",
- }
Genetic Programming entries for
DaeEun Kim
Citations
