Modèles graphiques stochastiques et optimisation pour la gestion des
réseaux écologiques
Nathalie Peyrard et Régis Sabbadin
Unité de Mathématiques et Informatique appliqués, Toulouse (MIAT)
INRA-ONIRIS,
Conservation of multiple species in food webs
On which species should we spend our money if our goal is to preserve biodiversity?
Conservation?
03/06/2015 INRA-ONIRIS, Nantes Régis Sabbadin
Conservation of multiple species in food webs
On which species should we spend our money if our goal is to preserve biodiversity?
Conservation?
03/06/2015 INRA-ONIRIS, Nantes
Spatial sampling of weeds for map reconstruction
How can we choose adaptively the locations to sample in
order to reconstruct a “reliable” weed map?
Collective management of crops resistance to pathogens
Where and when should we allocate
Resistant crops/protection actions?
These problems are ecological management problems
Choosing and applying conservation actions!
Conservation
These problems are ecological management problems
Choosing and applying conservation actions!
Choosing (adaptively) weed sample locations
These problems are ecological management problems
Choosing and applying conservation actions!
Choosing (adaptively) weed sample locations
Allocating crop systems in space/time
These problems involve networks
These problems involve uncertainty
Threatened species persistence is uncertain!
Weeds (especially the seed bank) are barely detectable!
Pathogens dynamics/spread are uncertain!
Artificial Intelligence tools for
ecological networks management
AI tools for the
Management of stochastic processes on networks
These tools are based on Stochastic graphical models
Bayesian networks
Markov Random Fields
(Factored) Markov Decision Processes
Plus the use of optimization/approximation methods
Dynamic programming
Reinforcement learning
Continuous optimization
Heuristics
Description of three applications
Applications in Ecology / Agro-ecology
Biodiversity conservation in Food webs
Optimal weeds sampling for map reconstruction
Management of ecological processes / crop pests
Application I : Conservation of multiple species in food webs
Problem : Multiple species protection, within a network of trophic relations
Management decision : which species should we protect (and when) , in order
Eve MacDonald-Madden & Iadine Chadès (CSIRO and University of Queensland) Peter Baxter, William Probert & Hugh Possingham (University of Queensland)
Edward Game (The Nature Conservancy) Nathalie Peyrard & Régis Sabbadin (INRA-MIAT)
Classically, trophic relations in food webs are:
deterministic, qualitative or quantified by mass flows (dynamical systems)
The Bayesian network approach is:
stochastic and quantified by conditional probabilities of presence
Food webs and bayesian networks
present/absent = 1/0
Absent Present
Present
P(HS,OF,CF) = P(HS|OF,CF) P(OF) P(CF)
1 1..
( ,..., n) i( | Pri ( ))i
i n
P S S P S eys S
=
=
∏
HS = « present/absent » (1/0)
P(HS=1|OF=0, CF=1)
Joint probability distribution over species occurrences:
« Conserving » a species increases its survival probability
And the survival probability of its predators
And thus, the species richness of the food web!
Find the « optimal » feasible set of species to conserve, given
A conservation budget B
Species conservation costs Ci
A global criterion (expectation of the number of surviving species)
Food webs, bayesian networks and
« optimal » conservation
Food webs, bayesian networks and
« optimal » conservation
0.686 1 0.1041
1 0.568
0.889 0.780
1 0.5601 0.858 0.566
0.621
0.2015
Problem: Optimal conservation is too difficult in general!
Food webs, bayesian networks and conservation heuristics
0 8 15 23 31 38 46 54
8 9 10 11 12
Expected number of species surviving
0 8 15 23 31 38 46 54
0.2 0.4 0.6 0.8 1
Percentage of budget required to manage all species Standard deviation
1
2
Budgeted Maximum Coverage Node Degree
Keystone Species Optimal
Betweeness Centrality BUP Rule
Greedy Heuristic
Structure : Alaskan Network
20 sets of randomly generated probability tables
Comparison of the expectation and variance of the number of surviving species
03/06/2015 INRA-ONIRIS, Nantes
Conclusions
A Bayesian network and combinatorial optimization model for species conservation within food webs
Efficient heuristics for approximating optimal conservation
No theoretical guarantee about the heuristics’ performance
Even more difficult in the “dynamic” case (work in progress)
Application II: Spatial weeds sampling for map reconstruction
Problem: an accurate map of weed repartition in a crop field
is a useful tool for studying weeds populations
but observations are costly
« Management » decision : where to get sample observations in order to achieve a good compromise between map accuracy and sampling cost?
Sabrina Gaba
(INRA- UMR Agroécologie) Mathieu Bonneau,
Nathalie Peyrard & Régis Sabbadin (INRA-MIAT)
Alexandre Albore
(ONERA-DCSD, INRA-MIAT) Florent Teichtel
(ONERA-DCSD)
Nathalie Peyrard & Régis Sabbadin (INRA-MIAT)
Optimal adaptive sampling strategy
MRF model
Reconstruction method
We combine MRF and MDP to model the problem of designing an adaptive strategy by optimization and we propose two heuristic solutions
Adaptive strategy
Estimated weed map
How to find the best one?
Performance of adaptive sampling heuristics
Comparison with 8 static sampling strategies
Z Reg1
Z
Reg2 Reg3
Star W1 W2
Performances of adaptive sampling heuristics
Benchmark of 6 real weed maps
Field of 13 by 13 quadrats, sample size = 13,5 % of total
NWC = number of well classified quadrats
Score = NWC(best strat) – NWC (strat)
Adaptive sampling is more efficient
Adaptive heuristics
Ongoing Work : Embedding adaptive sampling heuristics on a UAV
Mathematics and AI tools to adapt
sampling strategies to a robotic context
• ONERA/ région post-doc: Alexandre Albore (2014-2015)
• Co-advisor: F. Teichtel, ONERA
Objective:
To apply Artificial Intelligence methods for crops mapping using a UAV.
AI challenges
• Automated Planning under uncertainty: optimising the sampling strategy
Computer Vision: to identify and evaluate the collected data
Ongoing Work : Embedding adaptive sampling heuristics on a UAV
First step :
• Experiments on the MORSE simulator (LAAS-CNRS)
Conclusions
A framework and two heuristics for adaptive sampling under cost constraints
Adaptive sampling gives more accurate maps for the same cost
Method also applied on a problem of fire ants sampling
The sampled system is assumed static
Application III: Collective management of crop resistance to pathogens
Problem : Cultivar resistance enables to avoid fungicides but can be broken down
Management decision : Where and when should we allocate resistant crops to maintain both resistance and yield?
Benjamin Borgy (ex INRA-AGIR- MIAT) Jean-Noël Aubertot (INRA-AGIR)
Nathalie Peyrard & Régis Sabbadin (INRA-MIAT)
Crop rotation canola wheat barley
Host-pathogen interaction
+ -
avirulent
+ +
virulent
susceptible resistant
Case study: blackleg on canola
t, sowing t+1,sowing
harvest
tillage Dispersion from neighbouring wheat field
stubble
A GMDP model for the control of blackleg on canola
State variables in each field
crop
infection intensity
composition of pathogen population
Actions (on canola fields)
cultivar choice (CC)
ploughing threshold (PT)
Transition functions:
learned from simulations (SIPPOM )
Reward:
sum of local gross margins
Number of wheat neighbours> 1
Number of barley neighbors> 0
Number of barley neighbors> 3
CC=2 PT=1
CC=2 PT=4
CC=1 PT=1
CC=1 PT=4
0 1 2 3 4 5 6 7 8 9 10
0 5 10 15 20 25 30
0 10 20 30 40 50 60 70 80 90 100
0 5 10 15 20 25 30
27700 27800 27900 28000 28100 28200 28300 28400 28500
A tool for strategies simulation
Reward Infection intensity Proportion of virulents
Spatial repartition of actions
Conclusions
A framework for designing collective strategies by optimization
A model for comparing and simulating management strategies
BUT no satisfying solution to the problem of design by optimization
Ongoing Work :
Towards spatial management of weeds
How to optimise ecosystem services at the landscape scale?
• thesis of J. Radoszycki, ED Math/Info, 2012-2015
• Co-advisor: S. Gaba (UMR Agroécologie, Dijon)
+
-
Methodological tools
• Factored-Actions Factored MDPs (extension of GMDP)
• Stochastic factored policies
Bonus application:
Forest Management and risk aversion
(with S. Couture and M.-J. Cros)
03/06/2015 INRA-ONIRIS, Nantes
Multi-stand forest management under risk of windthrow
Forest owner with different multi-aged stand plots
Presence of windthrow risk
Risk-aversion of the owner -> Non-linear utility function :
A MDP framework for designing multi-stand strategies
Introduction of consumption-saving decisions in order to analyze the link between these decisions and forest management.
( )
1
( ) 1 U r r
β
β
=
−−
Forest Management and risk aversion
03/06/2015
Sensitivity analysis to
Storm risk
Risk aversion
Number of stands
N=5
N=10
Bonus application (2):
Leader-Follower MDP
Tomorrow’s talk by Anne-France Viet (MODEC)
General conclusions
Managing ecological networks
Networks can be: spatial, causal, …
Management can be: control, conservation, sampling ..
Ecosystems and agricultural systems : management share similarities
Common tools for all these problems
graphical models, simulation, optimization
Computing exactly the optimal strategy is out of reach
Current research focuses on approximate resolution
Still some challenges!
Design by optimization of strategies for managing ecological networks
How to improve solutions of problems with large factored state and action spaces?
We have heuristic solutions but there is still room for improvement. Planning as inference and continuous optimization are one possibility…
Dynamics
How to sample a system where the underlying process changes through time?
How to manage processes over an ill-know network?
Combine network learning and control/conservation actions optimization
Embedded systems
Using robots/UAV for spatial management/mapping
…
References
Food webs management
W.J.M. Probert, E. Mc Donald-Madden, N. Peyrard and R. Sabbadin. Computational issues surrounding the dynamic optimization of management of an ecological food web. ECAI 2012 Workshop AIGM.
E McDonald-Madden, R Sabbadin, P.W.J. Baxter, I Chadès, E.T. Game and H.P. Possingham. Using food webs to manage ecosystems, under review.
Adaptive spatial sampling
N Peyrard, R Sabbadin, D Spring, B Brook and R Mac Nally. Model-based adaptive spatial sampling for occurrence map construction. Statistics and Computing, 1-14, 2011.
M. Bonneau, N. Peyrard and R. Sabbadin. A Reinforcement-Learning Algorithm for Sampling Design in Markov Random Fields, ECAI 2012
M. Bonneau, S. Gaba, N. Peyrard and R. Sabbadin. Reinforcement learning-based design of sampling policies under cost constraints in Markov random fields: Application to weed map reconstruction, CSDA, 2014
A. Albore, N. Peyrard, R. Sabbadin and F. Teichteil. An Online Replanning Approach for Crop Fields Mapping with Autonomous UAVs, ICAPS, 2015.
GMDP, FA-FMDP, applications to agro-ecological processes management
N Peyrard, R Sabbadin, E Lo-Pelzer and J.N. Aubertot. A graph-based Markov decision process applied to the optimization of strategies for integrated management of diseases. Phytopatthology, 2007.
R Sabbadin, N Peyrard and N Forsell. A framework and a mean-field algorithm for the local control of spatial processes. International Journal of Approximate Reasoning, 2011.
J. Radoszycki, N. Peyrard and R. Sabbadin. Finding good stochastic factored policies for factored Markov decision processes, ECAI (poster), 2014.
The models
Models based on Stochastic graphical models
Bayesian networks
Markov Random Fields
(Factored) Markov Decision Processes
Bayesian networks
P(HS|OF,CF)
1 1..
( ,..., n) i( | Pri ( ))i
i n
P S S P S eys S
=
=
∏
Joint probability model:
Conditional probabilities
Concisely express joint probability distributions over sets of variables
A directed acyclic graph over variables
Conditional probability tables
Markov random fields
A framework for representing uncertain knowledge about spatial processes
Network representation of a spatial process
Undirected graph with cycles
A set of potential functions over cliques:
MRF probability distribution:
( )
c( )
cc C
P x x
∈
∝ ∏ Ψ
( ) 0,
c
x
cx
cΨ > ∀
c
( ) x
cΨ
Graph-based Markov Decision Processes
Several state variables {Xi}
Structured problems of sequential decision under uncertainty
Graph-based Markov Decision Processes
Several state variables {Xi}i∈V and decision variables {Di}i∈V
Structured problems of sequential decision under uncertainty
Graph-based Markov Decision Processes
Several state variables {Xi}i∈V and decision variables {Di}i∈V
A factored stochastic transitions model:
(
t 1 t, t)
i(
it 1 tN i( ), it)
i V
p x + x d p x + x d
∈
=
∏
Structured problems of sequential decision under uncertainty
Graph-based Markov Decision Processes
Several state variables {Xi}i∈V and decision variables {Ai}i∈V
A factored stochastic transitions model
A local reward model
(
t,
t,
t 1)
i(
it,
it,
it 1)
i V
r x d x
+r x d x
+∈
= ∑
Structured problems of sequential decision under uncertainty
Graph-based Markov Decision Processes
Optimization problem:
Find a policyδ : X → A assigning an action δ(x) to every possible states of the system, maximizing the expected discounted sum of future rewards
Local policies (approximate):