Holistic Modeling and Optimization

of Crowd Guidance in
Building Emergency Evacuation

Peng Wang, Peter B. Luh,

Shi-Chung Chang, Jin Sun
 Building Emergency Evacuation

l Building emergency evacuation

is of growing concern
– Emergencies include fire,
chemical spills, etc.
l Effective crowd guidance can
improve egress efficiency and
occupant survivability
l Existing guidance facilities
– Guidance: exit signs, audio instructions
– Static guidance versus dynamic fires and crowd movement

2 /28
 Building Emergency Evacuation
l To reduce the egress time the potential disasters such as
stampeding or blocking should prevented.
l Can guidance help?
– Traditional guidance is almost static, and does not consider
how hazard event dynamics affects people’s behavior and
cannot effectively prevent blockings in emergencies
– Our method:

3 /28
 Building Emergency Evacuation
l A model which can capture blockings is needed
– The model predicts the potential blockings in the future based
on current information of fires, egress and crowd movement
– Guidance is updated to mitigate or prevent blockings

4 /28
 Building Emergency Evacuation
l Existing models
– Crowd movement in
egress is captured by a
network-flow model
where stampeding or
blocking events cannot
be captured
– Helbing’s social-force
model captures blocking
events (faster-is-slower),
but focuses on one-room
scenarios, not in an
egress network

5 /28
 Difficulties in Modeling
l To capture blocking events it is necessary to combine
Helbing’s model with the network-flow model?
l Gap exists between the two models

Helbing’s model Network-flow model

- A microscopic model - A macroscopic model
- Equations for individual - Equations for collective
behaviors behaviors
- Individuals ↔ Particles - Crowd ↔ Flow

l To bridge the gap

– Crowd flow model is built up to translate Helbing’s microscopic
model to new macroscopic model (Presented last time)
6 /28
Table Contents

7 /28
A probabilistic graphical model
l The probabilistic graphical model
– Input: Guidance; Output: Crowd flow rate;
– Conditions: Fire status and path capacities
– Links: Conditional probability distribution

8 /28
 A probabilistic graphical model
l Faster-is-slower scenario
is achieved in this block
l If Qde  Ce , then Qe  Qde
with probability 1
l If Q de  Ce the probability
of blocking increases as
   
the difference of Qde and Ce
1  exp d  if Qe  Ce increases
d   Qe  Ce 
Pr(Qe | Qe , Ce )  
exp   if Q  Q Blc
  Qd  C  e e
  e e 

where   0
Question: Where is Qed from?

9 /28
A Probabilistic Graphical Model
l Conditional Probability Distribution Pr (Qed| we, sF)
The desired flow rate Qed(t): the number of people desiring to
move out during [ t , t  t ]
l If fire becomes closer, people become more impatient, and
the desired flow rate Qed increases in probability sense

10 /28
A probabilistic graphical model
xv: the number of
people in the area v

l Conditional Probability Distribution: Pr (we | ue, xv)

– Crowd response we: The number of occupant who will
follow the guidance at time [ t , t  t ]
– Suppose each individual will follow the guidance ue with
a certain probability (Trust Probability)
– The probability reflects people's inclination to use an
familiar exit

11 /28
A probabilistic graphical model
l The probabilistic graphical model incorporates two
psychology factors: impatience and trust
– Impatience is the cause of blocking events
– Trust on guidance reflects how guidance changes crowd
behaviors

12 /28
Table Contents

13 /28
 Egress Networks
l To incorporate the probabilistic graphical
model into the traditional network-flow
model let us review traditional network-
flow models first
l Review egress network
– Each area is represented by a node, and
the area capacity is ignored because the
bottleneck for crowd movement lies in the
path capacity
– Each path from one area to another is
represented by a directed arc with
specified capacity (persons per time unit)

14 /28
 Crowd Flow Equation
A simple example
l Review crowd flow equation
Q1 (t)
l The crowd flow equation is a linear x 1 (t) x 2 (t)

state equation Q3 (t)

– State: The number of people at each area Q2 (t)
x 3 (t)
x(t) = [x1(t), x2(t), x3(t)]'
– Flow: The movement of people over each arc
Q(t)  [Q1(t), Q2 (t), Q3 (t)]'
– Egress dynamics: Flow balance equation
 1  1 0 
 
x(t 1)  x(t) BQ(t) with egress matrix B   1 0  1
 0 1 1 

15 /28
Crowd Flow Equation
l Incorporate the probabilistic graphical model into the crowd flow
balance equation the new balance equation is given by
x(t  1)  x(t)  B Q (t)

x ( t  1)  x ( t )  B Q (u ( t ) | s F ( t ), C)

l The system dynamics

involves two stochastic
processes:
Fire process
Crowd movement process

16 /28
Table Contents

17 /28
The Constraints and Objective Function
l Constraints for guidance
– Never guide crowd to an area currently on fire or to be on
fire in near future.
– Guidance constraint is obtained based on prediction of fire
propagation
l The objective function to be maximized is
Cumulative number of people The total number of
evacuated during [0, T] people evacuated
J   t 1 t  ( x exit ( t  1)  x exit ( t ))  c T x exit (T )
T

 t 0 x exit ( t )  (c T  T) x exit (T)

T 1
xexit(t): The number of people in exit
areas at time t
cT: The weight for terminal time; cT>>T
18 /28
 The Objective Function
l Due to uncertainty both mean and semi-variance is calculated
where the semi-variance is a risk measurement
T 1 T 1
– Cumulative term: R 1   E[ x exit ( t )]  c ins  varsemi [ x exit ( t )]
t 0 t 0

– Terminal term: R 2  E[ x exit (T )]  c avg varsemi [ x exit (T )]

l Objective Function: Maximize J, with J  R 1  (c T  T )R 2

19 /28
Table Contents

20 /28
Solution Methodology
l In our problem the computation complexity is a challenge
– State space is large (the number of occupant in every area)
– The complexity is mainly from the huge state space
l To reduce the computation time Lagrangian Relaxation
(LR) is applied
– decompose overall
way-finding problem into
subproblems
by grouping
occupants

21 /28
 Solution Methodology - LR
l How Lagrangian Relaxation (LR) works in our problem
– The joint constraints is the path capacities shared by groups
– Decompose the overall problem by relaxing joint constraints
– Coordinate the solutions of the decomposed problems through
Lagrangian multipliers
Shared capacity Decomposition

x x

Lagrangian Multipers
(Coordination)
22 /28
Solution Methodology - LR
l Joint constraints exist in shared path capacities and are
embedded in the probabilistic graphical model.
l Approximation method is used to relax the joint constraint

23 /28
Solution Methodology - LR
l Approximation method is used to relax the joint constrain
of the path capacities
– First, the original holistic graphical model is separated by
making path capacities go to infinity for subproblems
– Second, a constraint is added to the outputs in the separate
graphical models.

~
Step 1: C e  
n
Step 2:  e ( t )]  C e
E[
i 1
Q (i )

24 /28
Solution Methodology - SDP
l As a result Lagrangian Relaxation can be applied and the
subproblem can be solved by the stochastic dynamic
programming (SDP) with time steps as stages
– SDP looks backward
– SDP yields a NP hard
problem
– In our problem the
computation complexity
is mainly from the large
size of state space
– SDP guarantee the
optimality

25 /28
Solution Methodology - Rollout

l Rollout algorithm with one-step look ahead policy

– Rollout algorithm looks forward from current state
– The states inaccessible from
the current state are not
included in computation,
thus the state space for
computation is reduced
– The tradeoff is the
optimality of the solution

26 /28
Solution Methodology - Rollout
l Rollout algorithm for a single group way-finding problem
* ~
u ( k )  arg max E{g k [ x ( k ), u ( k ), w ( k )]  Jk 1[f k [ x ( k ), u ( k ), w ( k )]]}
u(k)

– The heuristic policy:

Decision by heuristic
 Distance to Exit 
 arg min  
u
 Movement Speed 
 Distance to Exit 
 arg min  
u
 Flow Rate 

where flow rate is calculate

by the graphical model
27 /28
Table Contents

28 /28
Thank you