Case-Based Reasoning for Group Strategy and Coordination in

Multi-Robot Systems
Evan A. Krell
I. INTRODUCTION This review is concerned with how such group
experiences can be used for coordinated or group-
Case-based reasoning (CBR) is the process of encompassing problem solving. While artificial
using experiences as the basis for solving problems neural networks and other alternative machine
similar to those previously encountered. This has learning techniques have been demonstrated to be
been largely explored in cognitive studies and has effective at the individual level for perception
been used to improve fields such as health and (categorization) and parameter optimization, such
operations management. For example, an individuals techniques tend to be highly specific and difficult to
medical record can be compared to those in a generalize or transfer between robots. Symbol
database to diagnose diseases and prescribe grounding of learned concepts has even shown to be
treatments that were shown to be effective. The difficult on robots using the same sensors [Kira]. As
process has been formalized and applied to artificial it typically using the more general and
intelligence where problems are solved by using understandable case descriptions as local or global
similarity measures to compare a current problem to world states, case-based reasoning is potentially an
recorded scenarios in a database called the case base. effective approach to learning and problem solving at
Each scenario has a solution, and the most relevant the group level.
cases solution is chosen and adopted to suit the
current problem. More formally, case-based Research in CBR for multi-robot systems has
reasoning follows a 4-step cycle where the system (1) been active since the late 90s [source], but is still
retrieves similar case(s) using some measurement largely confined to the well-understood robot soccer
scheme. A solution is selected from the cases and (2) domain using teams of physically homogeneous
reused by adapting to the current problem. After robots. Of interest is where current knowledge of
attempting to apply the solution, it may be necessary CBR for MRS is at in relation to the goal of broader
to further (3) revise the solution to correctly fit the impact in diverse domains such as exploration and
problem. If the solution is successful, the problem disaster recovery. What challenges still need to be
and solution are (4) retained by adding to the case addressed to go from soccer to useful application?
base. Thus, the cycle concludes by learning through
new experiences. The review is organized by where the CBR is
performed in the group organization. It will begin by
In a multi-robot system (MRS), multiple and looking at cases where CBR was used in highly
possibly heterogeneous agents work together, either centralized environment where the CBR component
explicitly or implicitly, within the same domain and is used for mission planning and robot member
toward some common object. While individual group selection before the actual MRS is actually executing.
members may have particular roles and It then focuses on research where an online software
characteristics, research has been done to explore the centralized coordinator uses CBR to direct group
use of CBR for using overall group experiences for strategies and coordinated action sequences for the
overall strategy or multi-agent coordinated actions. team. Then, attention turns to instances where the
coordinator is an actual robot member of the
distributed system. Finally, cases are shown where
CBR is performed in a decentralized manner by
multiple individuals but for coordinated activity. Not
included are situations where CBR is used by an
individual robot only to optimize an individualistic
action such as finding the best angle to kick a ball
toward the goal. However, if the robot was using
CBR to determine that it should pass the ball or
otherwise respond to the group dynamics then it
would be relevant.
II. LITERATURE REVIEW mission planning system that uses CBR to analyze
mission objectives to generate mission plans and
In [], CBR is used for mission planning and select participants among a pool of available
selection of robot members for an MRS used for the heterogeneous robot platforms. Even though this
simulated naval missions of mine clearance, target research is not using CBR within the MRS itself, it
interception and tracking. The use of CBR is not demonstrates that past mission experience can be
performed by a member of the MRS, but rather a used to consider suitability of different types of
robots for mission strategies and takes advantage of CBR. First, the coach uses other machine learning
robot interactions. techniques to determine what aspects of a match were
most important to the match outcome. This is a
A similar approach is used in [], where a non- deeper level of case analysis that the typical metric
player software coordinator selects the team line up similarity approach. Second, the coach has measures
for a particular robot soccer match among a team of for similarity between heterogeneous robot types.
available heterogeneous players. The coach has a This is another example where awareness of robot
complete global view of the soccer environment with differences helps with effective MRS composition for
which to select players given a case base of past a particular problem.
soccer matches. The goal is to have the coach
dynamically switch out members during a game, but
currently the coach only selects before the match.
This paper makes two important contributions to III. CONCLUSION