Regularization multiplier

1,417 views
Skip to first unread message

Núria Roura

unread,
Feb 21, 2007, 9:45:14 AM2/21/07
to Maxent
Hi,

My name is Núria Roura, and I'm using Maxent to predict the
distribution of a species at regional scales. Since the area predicted
by Maxent models changes with the regularization value, I would like
to know which is the optimal values to develop the models? Can the
program select automatically the most suitable regularization value
for our model?

Thank you very much for your help. Bests,

Núria

Steven Phillips

unread,
Mar 7, 2007, 3:48:32 PM3/7/07
to Max...@googlegroups.com
Hi Nuria,

The regularization values were tuned to give good predictive
performance on a large collection of species from diverse regions,
described in Elith et al (2006), "Novel Methods Improve Prediction of
Species' Distributions from Occurrence Data", published in Ecography.
There is wide range of number of records in that data, and a fair
amount of diversity in the environmental data, so the default
regularization values should be reasonable for your data. If you're
doing work outside the scope of that paper, for example projecting to
different regions or climate conditions, you might need to make some
adjustments to the default regularization and feature types.

-- Steven

Núria Roura

unread,
Mar 12, 2007, 4:22:36 AM3/12/07
to Maxent
Hi Steven,

Thank you very much for you answer.

In fact, I'm modeling the distribution of a species in the Iberian
Peninsula. Therefore, since Elith et al. (2006) only modeled the
distribution of species in northern Europe, I'm not sure if the
default values will be suitable. What do you think? Should I make some
adjustments?

Thank you in advance,

Núria

Steven Phillips

unread,
Mar 12, 2007, 12:25:01 PM3/12/07
to Max...@googlegroups.com
Hi Nuria,

I expect that the default values will work fine in most regions. The
primary conditions where I'd think much about changing them:

1. If you have lots of spatial autocorrelation in the presence data
2. If you're training on one region / time and projecting to a
different geographic area or set of climate variables.

For 1, you may need to increase the regularization to reduce
over-fitting. The best approach to handle the second case is still
work in progress.

-- Steven

Núria Roura

unread,
Mar 14, 2007, 7:56:01 AM3/14/07
to Maxent
Hi Steven,

Thank you very much for the information and help. So ... for the
moment, I will use the default values.

Bests wishes,

Núria

Reply all
Reply to author
Forward
0 new messages