Modeling Complex Sociotechnical Systems:Systemic Risk and Emergence

Abhishek Sivaram, Vipul Mann, Laya Das, and Venkat Venkatasubramanian∗

CRIS Lab, Department of Chemical Engineering, Columbia University

COMPLEX RESILIENT

INTELLIGENT SYSTEMS

Complex Sociotechnical Systems and Emergence

Figure 1: SARS (2003), BP Deepwater Horizon Oil Spill (2010), Subprime Crisis (2008),and Northeast Blackout (2003)

I Complex adaptive systems engineeringI Need to go beyond analyzing them as

independent one-off accidentsI Common underlying patterns behind systemic

failuresI Need a unifying complex systems engineering

perspective of sociotechnical systemsI Need to recognize emergent phenomena and

understand the underlying mechanisms

I Failures (lessons) at all levelsI IndividualI CorporationI Corporate boardI Government: policies and regulationsI CommunityI National

I Teleo-Centric System Model for AnalyzingRisks and Threats(TeCSMART) (Venkatasubramanian andZhang, 2016) Figure 2: TeCSMART

Figure 3: Comparative Analysis

Causal Modeling of Process Systems

I How do we understand the causal links between different variables ina process?

I Creating causal graphs using entropic correlation in timeI Data-driven setting for identifying flow of information in a process systemI Captures higher order correlations between system variablesI Hierarchical strategy to estimate causal links for the plant-level operations

I Directed Graph As a Modeling Tool for Analyzing SystemicRisk in Process Systems (Suresh et al., 2019)

Figure 4: Tennessee Eastman Process, Plant-level causal model, Unit level causal models

Process Modeling from Data

I What do neural networks learn?

I Hidden representations of deepneural network towardsfunction approximation andclassification

I Deep nets, a few complex patterns

I Wide nets, a lot of simple patterns

I Black box models like neuralnetworks fail to explain the reasonfor their recommendation

I Model Hypothesis Generation usingGenetic Algorithm

I Mechanism identification usingGenetic Feature Extraction andStatistical Testing (GFEST)

Figure 5: Data, Deep NetworkRepresentation, Wide NetworkRepresentation

Figure 6: GFEST algorithm

ReferencesYu Luo, Garud Iyengar, and Venkat Venkatasubramanian. Social influence makes self-interested crowds smarter: An optimal control

perspective. IEEE Transactions on Computational Social Systems, 2018.

Resmi Suresh, Abhishek Sivaram, and Venkat Venkatasubramanian. A hierarchical approach for causal modeling of process systems.Computers & Chemical Engineering, 123:170–183, 2019.

Venkat Venkatasubramanian and Zhizun Zhang. Tecsmart: A hierarchical framework for modeling and analyzing systemic risk insociotechnical systems. AIChE Journal, 2016.

Multi-Agent Control with Soft Feedback

I How to coordinate multiple intelligent agents such that the crowd iscollectively wiser?

I Regulator’s dilemma: balancing between over- and under-regulationI Over-regulation hinders innovation, progress, and economic growthI Under-regulation results in safety threats and risk

The i -th agent can accept, reject, or partially accept the soft feedback u:

z+i = (1−βi)

(gi(zi)+ωi

)+βiu, βi ∈ [0, 1], ωi ∼ N (0, σω), u =

∑i

zin

I Social Influence Makes Self-Interested Crowds Smarter:An Optimal Control Perspective (Luo et al., 2018)

Time (second)0 50 100 150 200 250

Dec

isio

n er

ror

(kca

l)

-500

-400

-300

-200

-100

0

100

200

300

400

500

Time (second)0 50 100 150 200 250

Dec

isio

n er

ror

(kca

l)

-500

-400

-300

-200

-100

0

100

200

300

400

500

Time (second)0 50 100 150 200 250

Mea

n sq

uare

d er

ror

(kca

l2)

#104

0

1

2

3

4

5

6

7

8

9Open loop (observed)Open loop (predicted)Closed loop (observed)Closed loop (predicted)

Time (second)0 20 40 60 80 100 120

Mea

n sq

uare

d er

ror

(kca

l2)

#104

0

0.5

1

1.5

2

2.5

3Open loop (observed)Closed loop (observed)Closed loop with estimated beta profile (predicted)Closed loop with robust beta (predicted)Closed loop with optimal beta (predicted)Closed loop with optimal beta profile (predicted)

Figure 7: Experiment, System Identification, and Optimal Control Results

Agent Performance on a Network Topology

I Optimal communication architecturesI Particle Swarm Optimization as test bed

I High information transfer hinders explorationI Low information transfer hinders efficiencyI Robust topologies are generally not efficient

I Design guidelines to ensure efficient and robust networks

Figure 8: Design Spectrum, Performance Results

cris.cheme.columbia.edu Mar, 2020 venkat@columbia.edu