Distilling Free-Form Ecological Theory Using High Frequency Data

hostitchΤεχνίτη Νοημοσύνη και Ρομποτική

23 Οκτ 2013 (πριν από 3 χρόνια και 9 μήνες)

60 εμφανίσεις


Distilling Free
-
Form Ecological
Theory Using High Frequency
Data

Kevin Rose

Model Discovery


Applications/needs




Data sets large, difficult to manage


how can
we identify patterns amongst millions of data
points?


Complexity


can we visually identify all
relationships ourselves?


Weaknesses




Not question driven


Only identifies relationships, not causality


Hard to deal with hysteresis

Eureqa software


320 nm UV light transparency data


Lake Tahoe, CA/NV (MLTP site)

December 6
th
, 2006

320 nm UV light transparency data


Lake Tahoe, CA/NV

December 6
th
, 2006

Let’s work backwards

Let’s pretend we don’t know this

???

320 nm UV light transparency data


Lake Tahoe, CA/NV (MLTP site)

December 6
th
, 2006

This took my PC 26 seconds

Genetic algorithms


Symbolic regression

Hampen sø

total P (ug/L)

22.7

total N (ug/L)

580

Chl a (ug/L)

5.32

DOC (mg/L)

3.1

color (m
-
1)

0.48

residence time (years)

1.4

watershed size (km
2
)

9.2

Hampen sø

Hampen sø

Hampen sø

Hampen sø

The model

Input parameters (11)


Day of year

Temp at 0.5, 1, 2, 3, 4, 5, 7, 9m


PAR


Wind Speed


Hampen sø

The model

Surface water temperature

Temp difference
between 1m and 5m

Temp at 4m

Wind Speed

Important model component

metalimnion

Only important
when large

Epi
-

hypo

Diel fluctuations

Strength of stability

Input parameters (11)


Day of year

Temp at 0.5, 1, 2, 3, 4, 5, 7, 9m


PAR


Wind Speed


Hampen sø

Hampen sø

Hampen sø

Hampen sø

Hampen sø

Hampen sø

total P (ug/L)

69

total N (ug/L)

699

Chl a (ug/L)

30.3

DOC (mg/L)

12.8

color

6.5

residence time (years)

2.1

watershed size (km
2
)

1.2

22.7

580

5.32

3.1

0.48

1.4

9.2

Grib sø

Grib sø

Hampen sø

The model

Surface water temperature

Temp difference
between 1m and 5m

Temp at 4m

Wind Speed

Important model component

metalimnion

Only important
when large

Epi
-

hypo

Diel fluctuations

Strength of stability

Input parameters (11)


Day of year

Temp at 0.5, 1, 2, 3, 4, 5, 7, 9m


PAR


Wind Speed


8.0

Grib sø

Grib sø

Grib sø

Phycocyanin


Can we predict phycocyanin signal?


Phycocyanin

Chlorophyll


Input parameters:


Chlorophyll


DO


Air temp


Day of year


Temp at 0.0,0.5,1,1.5,2,3,4,5,6,7,8,10,11,12,
13,14,15,16,17,18,19m


Best model used chlorophyll, DO, temp at 0.5m


Some success


R2 = 0.78 over ~3 weeks.


What now


Ken Chiu & students (SUNY Binghamton)
currently developing GLEON version of
Eureqa


Issues with Eureqa


covariation,
independence, causality


Eureqa free to download, easy to use


recommended you give it a try


Next steps?