生態文化ニッチモデリングによる分布推定

生態文化ニッチモデリングによる遺物分布推定

Eco-‐cultural niche modelling

近　藤　康　久東京工業大学大学院情報理工学研究科日本学術振興会特別研究員(PD) 総括班/A01班/B02班研究協力者 [email protected][email protected]

1

交替劇B01/B02合同班会議 2013.7.22 明治大学中野キャンパス

遺跡と間違えましたが...

生態ニッチモデリング Ecological niche modelling

ニッチ (niche) とは •  日本語では「生態的地位」 •  “The environmental requirements (bio@c or abio@c) that need to be fulfilled for a popula@on to survive” (Peterson et al. 2011: 276)

•  生息場所（空間） •  資源利用パターン

2

Peterson et al. 2011

ISNN: 978-‐0-‐691-‐13688-‐2

ニッチの定義に２流派あり Grinnellian niche vs. Eltonian niche

•  Grinnellian niche (Grinnell 1917) Concept defined on the basis of environmental space of scenopoe@c (noninterac@ng and nonlinking) environmental variables and focused on geographic scales and requirements.

•  Eltonian niche (Elton 1927) Concept defined at small spa@al extents at which experimental manipula@ons are feasible, emphasizing the func@onal role of species in communi@es, and including models of resource consump@on and impacts.

(Peterson et al. 2011: 272–273)

3 ISBN: 978-‐0-‐691-‐13688-‐2

生態ニッチモデリングの定義 Defini@on of ecological niche modelling (ECNM)

“Es@ma@on of the different niches (fundamental, exis@ng, poten@al, occupied), par@cularly those defined using scenopoe@c [=noninterac@ng and nonlinking] condi@ons. In prac@ce, carried out via es@ma@on of abio@cally suitable condi@ons from observa@ons of the presence of a species.” (Peterson et al. 2011: 271) →既知の生息地と環境情報から，機械学習によって生物種のニッチを推定する手法

4 ISBN: 978-‐0-‐691-‐13688-‐2

入力変数：位置情報と環境情報 Model inputs: occurrence and environmental data

生物の位置 Occurrence [x, y]

生態ニッチモデリング Ecological niche modelling

生物群系（植生） Biome

気候指標（気温&降水量） Clima@c indices

標高由来の地形指標 DEM-‐based topological indices

5

ヒメゴジュウカラの生態ニッチ Si#a pygmaea predic@ons in North America

6

生態ニッチモデリングの二大アルゴリズムTwo major algorithms of ecological niche modelling

GARP MaxEnt

遺伝的アルゴリズム

アルゴリズム algorithm

最大エントロピーモデル

在のみ presence-‐only

サンプリング sampling

在のみ presence-‐only

二値 [0, 1] binary

出力 output

連続的 [0...1] con@nuous

強い calibrated

バイアス補正 biased data

弱い biased

7 どちらもオープンソースのソフトウェアを入手可能

遺伝的アルゴリズム法

Gene@c Algorithm for Rule-‐Set Produc@on •  “A machine-‐learning gene@c algorithm originally developed for determining the ecological niches of plant and animal species” (Stockwell & Peters 1999)

•  Desktop GARP (open source soeware package) hfp://www.nhm.ku.edu/desktopgarp/

•  Also included in OpenModeller hfp://openmodeller.sourceforge.net

8

遺伝的アルゴリズムの実際 Stockwell 1999 in Machine Learning Methods for Ecological Applica9ons

1.  Start at ini@al @me t = 0 2.  Ini@alize a (usually random) popula@on of individuals P(t) 3.  Evaluate the fitness of all individuals 4.  Test for termina@on criteria, while not done do: 5.  Increase @me counter 6.  Create a new set of individuals P’(t) 7.  Recombine the “genes” of selected individuals

using heuris@c operators 8.  Evaluate new fitness of the popula@on P’(t) 9.  Select the survivors 10. Go to 4

9 ISBN: 978-‐0-‐412-‐84190-‐3

GARP parameter seings

•  1000 replicate runs •  0.01 convergence limit •  80% for training •  Hard, 0% omission error (i.e., failure to predict a known presence)

•  50% commission error (i.e., areas of absence that are incorrectly predicted present)

•  10 final “best-‐subset” models for each complex

10 hfp://www.nhm.ku.edu/desktopgarp/

最大エントロピー法

Maximum Entropy (Jaynes 1957) The best approxima@on is to ensure that: 1.  The approxima@on sa@sfies any constraints on

the unknown distribu@on that we are aware of; 2.  Subject to those constraints, the distribu@on

should have maximum entropy 11

Phillips et al. 2006

hfp://dx.doi.org/10.1016/j.ecolmodel.2005.03.026

最大エントロピー原理 Principle of maximum entropy

12

H (π̂ ) = − π̂ (x)ln π̂ (x)x∈X∑

a finite set of points

approxima@on of π (unknown probability distribu@on) to point x

entropy


最大エントロピーの推定 Maximum Entropy approxima@on (Phillips et al. 2006: 236)

•  Formalize the constraints on the unknown probability distribution π.

•  Assume a set of known real-valued functions f1 … fn on X, known as “features” [=environmental variables].

•  Assume that the information we know about is characterized by the expectations (averages) of the features under π.

•  Each feature fj assigns a real value fj(x) to each point x in X. •  The expectation of the feature fj under π is

13

π f j!" #$= π (x) f j (x)x∈X∑ (1)



•  π [fj] can be approximated using a set of sample points x1 … xm drawn independently from X.

•  The empirical average of fj is

14

π f j!" #$=1m

fj (xi )i=1

m∑ (2)

an estimate of π [fj]

•  By the maximum entropy principle, seek the probability distribution of subject to the constraint that each has the same mean under as observed empirically:

π̂π̂

π̂ f j!" #$= π f j!" #$ for each feature fj (3)



•  has an alternative characterization. •  Convex duality (Delle Pietra et al. 1997) shows that is

exactly equal to the Gibbs probability distribution qλ that maximizes the likelihood of the m sample points:

15

qλ (x) =eλ⋅ f (x )

Zλ

(4)

where •  λ is a vector of n real-valued coefficients (feature weights). •  f denotes the vector of all n features. •  Zλ is a normalizing constant to ensure qλ sums to 1.

π̂π̂



•  Equivalently, qλ minimizes the negative log likelihood of the sample points (“log loss”)

16

π[− ln(qλ )]= lnZλ −1m

λ ⋅ f (xi )i=1

m∑ (5)

•  Maxent tends to over-fit the training data. •  Therefore, the means under should be close to their

empirical values by relaxing the constraint in (3) π̂

π̂ f j!" #$− π f j!" #$ ≤ β j for each feature fj (6)

for some constants βj.



•  The Maxent distribution can now be shown to be the Gibbs distribution that minimizes

17

π[− ln(qλ )]+ β j λ jj∑ (7)

where •  The first term is the log loss (5)

•  The second term penalizes the use of large values for the weights λj.


環境変数 f の調整法は５通り Five feature types

1.  Linear feature

2.  Quadratic feature

3.  Product feature [for two variables]

4.  Threshold feature [not used]

5.  Binary feature [for categorical values v1 … vk] ith feature is 1 wherever the variable equals vi, otherwise 0.

18

π̂[ f ]

π̂[ f 2 ] − π̂[ f ]2

π̂[ fg] − π̂[ f ]π̂[g]10

!"#

if above a given threshold


MaxEnt parameter seings

19 hfp://www.cs.princeton.edu/~schapire/maxent/

生態文化ニッチモデリング Eco-‐cultural niche modelling (ECNM): an applica@on to human

•  過去の人類の生活が自然環境の影響を大きく受けたという前提に立てば，先史人類にも生態ニッチモデリングを適用できる。

•  人類は自然環境に対して他の生物とは異なるふるまいをする。 •  自然環境を克服するために，技術を容易に進化させる •  自然環境を改変する（農耕，家畜化，森林伐採etc.）

→ 生態ニッチをみずから変える能力をもっている → これこそが文化 (culture) の発現

20

入力変数：遺跡情報と古環境情報 Model inputs: archaeological and palaeoenvironmental data

遺跡の位置 archaeological sites [x, y]

生態文化ニッチモデリング Eco-‐cultural niche modelling

生物群系（植生） biome

古気候指標（気温&降水量） palaeoclima@c indices

標高由来の地形指標 DEM-‐based topological indices

21

交替劇のECNMは仏チームが先行 Preceding studies by Banks et al. (2008b, 2013)

22

Humaneclimate interaction during the Early Upper Paleolithic: testing thehypothesis of an adaptive shift between the Proto-Aurignacian and the EarlyAurignacian

William E. Banks a,b,*, Francesco d’Errico a,c, João Zilhão d

aCNRS, UMR 5199-PACEA, Université Bordeaux 1, Bâtiment B18, Avenue des Facultés, 33405 Talence, FrancebBiodiversity Institute, University of Kansas, 1345 Jayhawk Blvd, Dyche Hall, Lawrence, KS 66045-7562, USAcDepartment of Archaeology, History, Cultural Studies and Religion, University of Bergen, Øysteinsgate 3, 5007 Bergen, NorwaydUniversity of Barcelona, Faculty of Geography and History, Department of Prehistory, Ancient History, and Archaeology, C/Montalegre 6, 08001 Barcelona, Spain

a r t i c l e i n f o

Article history:Received 9 March 2012Accepted 4 October 2012Available online 13 December 2012

Keywords:Eco-cultural niche modelingProto-AurignacianEcological niche expansionHeinrich Stadial 4

a b s t r a c t

The Aurignacian technocomplex comprises a succession of culturally distinct phases. Between its firsttwo subdivisions, the Proto-Aurignacian and the Early Aurignacian, we see a shift from single to separatereduction sequences for blade and bladelet production, the appearance of split-based antler points, anda number of other changes in stone tool typology and technology as well as in symbolic material culture.Bayesian modeling of available 14C determinations, conducted within the framework of this study,indicates that these material culture changes are coincident with abrupt and marked climatic changes.The Proto-Aurignacian occurs during an interval (ca. 41.5e39.9 k cal BP) of relative climatic amelioration,Greenland Interstadials (GI) 10 and 9, punctuated by a short cold stadial. The Early Aurignacian (ca. 39.8e37.9 k cal BP) predominantly falls within the climatic phase known as Heinrich Stadial (HS) 4, and itsend overlaps with the beginning of GI 8, the former being predominantly characterized by cold and dryconditions across the European continent.

We use eco-cultural niche modeling to quantitatively evaluate whether these shifts in material cultureare correlated with environmental variability and, if so, whether the ecological niches exploited byhuman populations shifted accordingly. We employ genetic algorithm (GARP) and maximum entropy(Maxent) techniques to estimate the ecological niches exploited by humans (i.e., eco-cultural niches)during these two phases of the Aurignacian. Partial receiver operating characteristic analyses are used toevaluate niche variability between the two phases.

Results indicate that the changes in material culture between the Proto-Aurignacian and the EarlyAurignacian are associated with an expansion of the ecological niche. These shifts in both the eco-cultural niche and material culture are interpreted to represent an adaptive response to the relativedeterioration of environmental conditions at the onset of HS4.

! 2012 Elsevier Ltd. All rights reserved.

Introduction

The Proto-Aurignacian and the Early Aurignacian are chrono-logically and techno-typologically different phases of the Auri-gnacian cultural tradition. Between the two, we observe a change inreduction sequences used to produce blades and bladelets, theappearance of split-based antler points in the Early Aurignaciantoolkit, and a number of changes in lithic typology as well as insymbolic material culture (e.g., Conard, 2003; Bon, 2006; Liolios,

2006; Vanhaeren and d’Errico, 2006; Teyssandier, 2007; Zilhão,2007; Teyssandier and Liolios, 2008; Teyssandier et al., 2010).

The Aurignacian traditionally has been viewed as thecultural technocomplex associated with the movement of anatom-ically modern humans into the European continent and the subse-quent replacement of autochthonous Neanderthal populations.Teyssandier (2008) points out that, because of this viewpoint,technological and behavioral variability within the Aurignaciantended to be overlooked, and it was viewed as a homogenouscultural tradition. More recently, the situation has changed andAurignacian diversity has become a central subject of study (e.g.,Bon, 2002; Zilhão and d’Errico, 2003; Bar-Yosef and Zilhão, 2006;Pesesse, 2008; Michel, 2010). Such research is aided by the fact that

* Corresponding author.E-mail address: [email protected] (W.E. Banks).

Contents lists available at SciVerse ScienceDirect

Journal of Human Evolution

journal homepage: www.elsevier .com/locate/ jhevol

0047-2484/$ e see front matter ! 2012 Elsevier Ltd. All rights reserved.http://dx.doi.org/10.1016/j.jhevol.2012.10.001

Journal of Human Evolution 64 (2013) 39e55

PLoS ONE 3/2: e3972

64: 39-‐55

Banks et al. 2008: Neanderthal ex@nc@on by compe@@ve exclusion

23

Predic@ng the habitat of hunter-‐gatherers •  Who? …Neanderthals vs. AMHs •  Where? … Europe •  When? … Three sub-‐stages in MIS 3

•  How? … ECNM by GARP

Pre-‐Heinrich event 4 (Pre-‐H4)

Heinrich event 4 (H4)

Greenland Interstadial 8 (GI8)

43.3 – 40.2 ka 40.2 – 38.6 ka 38.6 – 36.5 ka

hfp://dx.doi.org/10.1371/journal.pone.0003972

Archaeological data: Pre-‐H4

24

Archaeological data: H4

25

Archaeological data: GI8

26

Palaeoenvironmental variables

27

•  Landscape aspects (slope, aspect, and topographic index) from Hydro-‐1K (USGS)

•  Clima@c simula@on by LMDZ3.3 Atmospheric Gerenal Circula@on Model (50km final resolu@on) •  Ice-‐sheet: ICE-‐4G reconstruc@ons for 14 kyr BP (Pel@er 1994) •  SSTs for the three sub-‐stages •  GARP-‐based simula@ons of

–  The coldest month/warmest month/annual temperature –  Precipita@on

•  GARP parameters are the same as the previous study


Palaeoenvironmental variables

28

Warmest month temperature Coldest month temperature

Mean annual precipita@on (mm x100) Mean annual temperature


Results: Neanderthal vs. AMH

29

Pre-‐H4 (43.3 – 40.2 ka)



30

H4 (40.2 – 38.6 ka)



31

GI8 (38.6 – 36.5 ka)


Counterfactual results for Neanderthals

32

If the Neanderthals survived as they did during H4 in the GI8 clima@c condi@ons, their niche would have been like this…


まとめ：生態文化ニッチモデリングの特徴 ECNM: Data-‐driven simula@on of the human past

•  データ駆動型のシミュレーション •  遺跡の分布のバイアスを低減できる。むしろ，バイアスも教師データとして分布予測に活用できる。

•  各環境変数の寄与度が評価尺度として重要。 •  静態的な分布を予測するツールなので，時系列のような動態分析には不向き。

•  人類進化の「なぜ」「どのように」に直接答えるものではないが，問題発見のためのツールと位置づけることができる。

33

生態文化ニッチモデリングによる分布推定

Technology

Transcript of 生態文化ニッチモデリングによる分布推定