Feedback on Reviews of Leow Paper.

FLAWS: There are a number of criticisms you could make of this
paper.

1. There is no justification of the use of a simple inverse square law
odor distribution. Real odor distributions tend to form into plumes -
with a gaussian distribution giving a normal (bell-shaped) curve -
rather than diffuse evenly in all direction. The inverse-square
distribution is actually a better model of sound or light than
odor. In which case, is the paper really about *sound*-based
navigation? Further, does his model of odor distribution around
obstacles fit reality?

2. The model is a 2D one, whereas most creatures live in a 3D
one. In particular, this calls into question the conclusion about
the need for at most 2 sensors. 

3. The null hypothesis is not tested, ie would these models cope
as well even if there was no odor present? In his defence, one might say
that the null hypothesis is too obviously false to be worth
testing, but sometimes the obviously false turns out to be true.

4. The investigation of the noisy environment set randomness to its
"best" value for each of the networks, except for the spatial
difference network, which was given a parameter value of 0.01, instead
of 0.3-0.4. No explanation of this was given.

5. It has been noted that similar paper by the same author had
previously appeared: Wee Kheng Loew and Hsueh Chyi Wee "A minimal
neural network for target searching and danger avoidance by smell", in
Proceedings of International Conference on Neural Networks, vol. I,
580-585, 1997. It is accepted practice in Informatics to republish
extended and more mature conference papers in a journal. This is to
allow fast publication of preliminary versions of the research, and
then make a fuller archival journal publication, without the
conference's space and time constraints. However, one would normally
cite the conference paper in the journal paper, mention that the
journal paper is an extended version of the conference one and give
the two papers the same name to avoid one's publication list giving a
misleading impression of the number of distinct publications. Leow
doesn't do any of these.

6. There were papers on modelling smell prior to 1998 that were not
referenced by the authors. I would not expect you to know this,
although it is possible to track these down.

7. The paper has a number of minor spelling and grammatical
errors. The author is probably not a native English speaker, but
he should have sought help from someone who was.  

HYPOTHESES: There were a number of more or less implicit
hypotheses in the paper and one clearly explicit one. The
experimentation was mainly exploratory: trying to form and then
confirm some hypotheses. The hypotheses were probably anticipated
in advance to varying degrees, but the most explicit one (3
below) appears not to have been anticipated.

1. That the various neural nets modelled real animals' abilities to
use smell sensing, ie the model's behaviour was appropriate over
a range of experimental situations. 

2. That the quality of the behaviour varied monotonically with
the complexity of the models. 

3. However, the behavioual differences in the 3 most complex
models was not significantly different, leading to the claim
that 2 smell sensors was sufficient.

4. That all 4 models were affected by the degree of wandering and
by noise, but that only the temporal difference model was
significantly affected.

5. That increasing hunger makes the models take more risks. 

OTHER POINTS: 

1. This paper motivated a fairly new problem of modelling smell
sensing, applied an existing technique of neural nets to this
problem, produced a model of a natural system, then conducted an
exploratory study to identify some hypotheses. The explicit
hypothesis that most people identified does not appear to have
been anticipated, so this was not a study to confirm a
pre-existing hypothesis as far as claim 3 (and 4?) goes (although
one might think it was for the implicit 1 & 2. Nor was there really
a contribution of improving, extending or adapting neural nets,
although some people listed this.

2. It is quite standard for the abstract to adapt or repeat text
from the introduction (or conclusion), so criticising this was a
bit unfair.

3. There are two kinds of discussion of related work: 
	* the literature survey, which normally comes early in
	  the paper and is part of the motivation: it is a broad
	  and shallow survey of the field.
	* the related work, which normally comes late in the
          paper and is part of the evaluation: it is a narrow and
	  deep comparison of the author(s)'s work with that of
	  rival approaches.
Some people confused these and thought the comparison in the
discussion section should have been in the introduction. 

Having said that, the author doesn't really go very much deeper in the
later section than in the earlier one.

5. Although the models were implemented with neural nets, no
machine learning was involved in the project. The nets appear to
have been hand-coded and were not modified during the
experiments. So rival approaches were *not* other machine
learning techniques, but other modelling techniques and alternate
models.