Feature Selection for CBR in Imitation of

RoboCup Agents: A Comparative Study

Edgar Acosta, Michael Floyd, and Babak Esfandiari
Network Management and Artificial Intelligence Laboratory
Carleton University, Ottawa, Canada
{acosta.edgar,mfloyd,babak}@nmai.ca
http://www.nmai.ca/research-projects/agent-imitation

Abstract. We present a comparison of two methods for selecting the

features employed in the representation of cases for a Case Based Reasoning approach to imitation of agents in the RoboCup simulation platform.
Feature Selection methods are compared empirically in terms of the imitation performance, the size of the resulting feature set, and execution
time. Differences and their implications are discussed.

Introduction

One central issue for every artificial intelligence problem is that of knowledge
representation. Decisions regarding to what to represent and how to do it often
facilitate, and likewise constrain, the problem solving process and its success.
The most sensible way to deal with such sensitive decisions involves getting
expert advice. However, when there is no expert knowledge available, either
because of the novelty of the problem or due to scarcity of resources, feature
selection methods can help to identify the pieces of information that are more
relevant in solving the task [1].
Another aspect in which expert knowledge eases the solution of an artificial intelligence problem regards the processes underlying the task itself. When
dealing with a new domain, or when the mechanisms leading to the problem
solution are not well known, Case-based reasoning provides a powerful learning
methodology [2].
Imitation is one such task, in which there is no formal model of the observed
behavior. The RoboCup Simulation Imitation project [3] employs both casebased reasoning and feature selection on learning to reproduce the behavior of
soccer playing agents based on the agent inputs, which represent an incomplete
view of the soccer field.
We compare two feature selection methodologies in order to evaluate which
methodology would better help on overcoming some of the limitations of the
current case recognition framework of the RoboCup Simulation Imitation.

Edgar Acosta, Michael Floyd, and Babak Esfandiari

The RoboCup Simulation Imitation Project

Imitating and agent involves discovering both a model of the environment and a
model of the behavior of the target agent as it interacts with the environment.
The environment model consists of representations of objects and events occurring in the environment. The behavior model, on the other hand, consists of
an interpretation of the environment, including a set of actions that the agent
performs upon the occurrence of events in the environment.
Although our aim is to investigate imitation in independence of the domain,
the RoboCup Simulation League provides an exceptional testbed for our project.
Imitation experiments in the context of the RoboCup Simulation League are
facilitated by a number of factors:
availability of an open source RoboCup Simulation server,
availability of a growing number of RoboCup agents of diverse complexity,
easiness to gather data at any time.
As the results of our imitation efforts improve, well try to imitate in other
domains.
2.1

The RoboCup Simulation League

In the RoboCup Simulation League a simulation server provides software agents

with limited and noisy information about objects and agents in a virtual soccer
field. In a RoboCup competition, agents work in a soccer team trying to score
and to stop the efforts of the opponent team.
Every 100 ms. the simulation server sends messages to each agent regarding
the position, distance and movement of visible objects in the visual field of
the agent (see messages), aural events in the aural field of the agent (hear
messages), the physical condition of the agent (sense body messages), and other
information such as the score.
Agents have less than 100 ms to send commands to the server. Commands
regard actions performed attempted by the agent such as kick, dash, turn, and
say.
2.2

The Case Based Reasoning Approach

Case Representation what a case represents what is encoded in a case what

are the parameters of the CBR and the case search.
Limitations problems with the case representation problems with the search
Improving the Case Representation

Feature Selection for CBR in Imitation Agents: Comparative Study

Feature Selection Methods for CBR

Rough Set Theory Approach to Feature Selection

4.1

Benefits of Rough Set Theory Methods

4.2

Proposed Feature Selection Method

4.3

Hypotheses

5
5.1

Feature Selection Methods Comparison

Materials

Target Agents
Independent variables
Dependent variables
5.2

Methodology

Imitation Performance metrics

5.3

Results

Imitation Performance
Feature Set Size
Feature Selection Execution Time

Discussion

6.1

Implications for the Case Representation

6.2

Implications for the CBR based Imitation

6.3

Increasing the Feature base

Conclusion

References
1. Liu, H., Motoda, H.: Less is more. In: Computational Methods of Feature Selection.
Chapman & Hall/CRC, Boca Raton (2008)
2. Pal, S.K., Shiu, S.C.K.: Introduction. In: Foundations of soft case-based reasoning.
Wiley-Interscience, Hoboken, N.J. (2004)
3. Floyd, M.W., Esfandiari, B., Lam, K.: A case-based reasoning approach to imitating
robocup players. In: Proceedings of the Twenty-first International Florida Artificial
Intelligence Research Society Conference, FLAIRS 2008, AAAI Press (2008) 251
256