Newsgroups: comp.ai
Path: cantaloupe.srv.cs.cmu.edu!das-news2.harvard.edu!news2.near.net!news.mathworks.com!usenet.eel.ufl.edu!news-feed-1.peachnet.edu!hobbes.cc.uga.edu!cssun.mathcs.emory.edu!emory!swrinde!pipex!uknet!festival!toby
From: toby@castle.ed.ac.uk (R T Tyrrell)
Subject: how to choose behaviours
X-Newsreader: TIN [version 1.1 PL8]
Message-ID: <D6ArFM.9CC@festival.ed.ac.uk>
Sender: news@festival.ed.ac.uk (remote news read deamon)
Organization: Edinburgh University
Date: Fri, 31 Mar 1995 08:54:58 GMT
Lines: 121



  3 papers on behavioural choice (action selection) are available by ftp:

    1. An evaluation of Maes' "bottom-up mechanism for action selection".

    2. The use of hierarchies for action selection.

    3. Defining the action selection problem.


Ftp details are given below, followed by abstracts.  The papers consider the
problem of how best to choose one out of a set of candidate behaviours,
given certain needs on the part of the agent, and certain dangers/
opportunities present in the environment.  Analysis of this problem is
presented, together with discussion and evaluation of different mechanisms
which attempt to solve the problem (the mechanisms come from the fields of
both animal behaviour and robotics).  The papers relate to topics such as
motivations, decision-making, goal-seeking, drives, and reactive decisions
vs. traditional planning.


Toby Tyrrell
Plymouth Marine Laboratory,
Prospect Place,
Plymouth,
PL1 3DH
UK.

email: tt@uk.ac.pml

--------------

To get one of the papers:

1.   'ftp octopus.npm.ac.uk'  (user: anonymous  password: <userid>)

2.   'cd /pub/srp1'

3.   'binary'

4.   'get tt1.ps.Z'  (or tt2.ps.Z, tt3.ps.Z)

5.   'quit'

6.   'uncompress *.Z'  then print or view

--------------

An evaluation of Maes' "bottom-up mechanism for action selection".

Abstract:
Maes has proposed a mechanism for action selection (behavioural choice)
[Maes, 89; 90; 91a].  This mechanism is reviewed here and is evaluated
using a simulated environment.  The simulated environment is a detailed and
complex generalised model of the action selection problem faced by an animal
in the wild, and presents a severe test for an action selection mechanism.
The results of testing Maes' mechanism in the simulated environment are
discussed , some observed deficiencies in the mechanism's operation are
described, and the computational reasons underlying the deficiencies are
explained.  It is argued that some central aspects of the design of Maes'
mechanism mean that it is not well able to deal with animal-like action
selection problems.



The use of hierarchies for action selection.

Abstract:
Several researchers of animal behaviour, such as Tinbergen and Baerends,
have proposed hierarchical mechanisms for action selection.  Maes, amongst
others, has argued against mechanisms of this type because of the top-down
control and lack of robustness inherent in these {\it hierarchical decision
structures} (rigid switching mechanisms) in which decisions are made at
every level.  Two alternatives exist to these hierarchical decision
structures: {\bf (1)} Maes and others have designed non-hierarchical action
selection mechanisms, and {\bf (2)} Rosenblatt \& Payton have outlined a 
hierarchical mechanism which does {\it not} make decisions at every level.
In this paper the use of the two types of hierarchies for action selection
is discussed and the Rosenlatt \& Payton approach with free flow of
information, combination of evidence and the ability to select compromise
candidates is supported.  It is also argued that the problem of action
selection is by nature intrinsically hierarchical, and so Rosenblatt \&
Payton-like hierarchies ({\it free-flow hierarchies}) are more suitable for
action selection than non-hierarchical mechanisms.  Evidence from
observation of animal behaviour and from computer simulation testing is used
to give support to these claims.



Defining the action selection problem.

Abstract:
There has been a lack of progress in the field of action selection due to
an incomplete understanding of the problem being faced.  The differing
nature of constituent parts of the action selection/`time-allocation'
problem has not been properly appreciated.  Some common sub-problems, such
as obtaining food and avoiding predators, are described in terms of the
demands they make on an animal's time.  The significant differences between
these sub-problems are highlighted and a {\it classificatory scheme} is
proposed, with which sub-problems can be categorised.  The need to take into
account the full range of different sub-problems is demonstrated with a few
examples.  A particular shortcoming shared by all of the more well-known
action selection mechanisms, from both robotics and animal behaviour, is
described.








--

Toby Tyrrell
Plymouth Marine Laboratory,
Prospect Place,
Plymouth PL1 3DU.

tt@pml.ac.uk
