You are on page 1of 12

Mathematical Programming Framework for UE

• Solution to Wardrop’s User Equilibrium Problem is obtained by formulating the


following equivalent problem:
𝑉𝑎
min 𝑍 𝑇𝑖𝑗𝑟 = ෍ න 𝐶𝑎 𝑣 𝑑𝑣
𝑎 0

• Subject to
෍ 𝑇𝑖𝑗𝑟 = 𝑇𝑖𝑗
𝑟
and
𝑇𝑖𝑗𝑟 ≥ 0

The objective function is the sum of the areas under the cost-flow curves for all
links in the network
Solution Algorithm for UE
• The Lagrangian for the UE minimisation problem can be written as:

𝐿 𝑇𝑖𝑗𝑟 , 𝜆𝑖𝑗 = 𝑍 𝑇𝑖𝑗𝑟 + ෍ 𝜆𝑖𝑗 𝑇𝑖𝑗 − ෍ 𝑇𝑖𝑗𝑟


𝑖𝑗 𝑟
Where, 𝜆𝑖𝑗 is Lagrange multiplier.
The solution to the UE problem is obtained by minimising the Lagrangian 𝐿 𝑇𝑖𝑗𝑟 , 𝜆𝑖𝑗 .
Frank – Wolfe (convex combinations) algorithm is used to solve the above objective
function.
In this algorithm, an equivalent linear-objective-function-solution obtained in each
iteration is used to set the flows for the next iteration.
The linear objective function that is minimised in each iteration is:

𝑍 𝑛 𝐹 = ෍ 𝐶𝑎𝑛 𝐹𝑎 where, 𝐶𝑎𝑛 = link cost in the 𝑛𝑡ℎ iteration and 𝐹𝑎 is auxilary flow on link 𝑎
𝑎
This calls for minimising the total travel time over the network for fixed costs, 𝐶𝑎𝑛 . This
is achieved by an all or nothing assignment.
Solution Algorithm for UE

• The current flows are then calculated as:


𝑉𝑎𝑛+1 = 1 − 𝜙 𝑉𝑎𝑛 + 𝜙𝐹𝑎 , with 0 ≤ 𝜙 ≤ 1
Where, 𝑉𝑎𝑛 =current flow on link a and 𝜙 is chosen in such a way that the
following objective function is minimised:
1−𝜙 𝑉𝑎𝑛 +𝜙𝐹𝑎

min 𝑍 = ෍ න 𝐶𝑎 𝑣 𝑑𝑣
𝑎 0
Thus, the optimum value of 𝜙 is obtained by solving
𝜕𝑍
= (𝐹𝑎 − 𝑉𝑎𝑛 ) 𝐶𝑎 1 − 𝜙 𝑉𝑎𝑛 + 𝜙𝐹𝑎 = 0
𝜕𝜙
Frank-Wolfe Solution Algorithm for UE
• Step 0: Initialisation
Set all initial link flows to 0, and perform all or nothing assignment based on 𝐶𝑎 = 𝐶𝑎 (0) for all a. This yields link
flows 𝑉𝑎1 . Set counter n: = 1
• Step 1: Up date.
Set 𝐶𝑎𝑛 = 𝐶𝑎 (𝑉𝑎𝑛 ) for all a

• Step 2: Direction finding.


Perform all-or-nothing assignment based on {𝑉𝑎𝑛 }. This yields a set of (auxiliary) flows {Fa}.
• Step 3: Line search.
Find 𝜙 that solves
1−𝜙 𝑉𝑎𝑛 +𝜙𝐹𝑎

min 𝑍 = ෍ න 𝐶𝑎 𝑣 𝑑𝑣 i.e., by solving σ𝑎(𝐹𝑎 − 𝑉𝑎𝑛 ) 𝐶𝑎𝑛 1 − 𝜙 𝑉𝑎𝑛 + 𝜙𝐹𝑎 = 0


𝑎 0
• Step 4: Move.
Set 𝑉𝑎𝑛+1 = 1 − 𝜙 𝑉𝑎𝑛 + 𝜙𝐹𝑎 , with 0 ≤ 𝜙 ≤ 1

• Step 5: Convergence test.


If a convergence criterion is met, stop (the current solution, {𝑉𝑎𝑛+1}, is the set of equilibrium link flows);
otherwise, set n = n + 1 and go to step 1
2000

1000

0 Bisection method can be used for finding


0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1
the value of 𝜙 in the equation,
-1000 σ𝑎(𝐹𝑎 − 𝑉𝑎𝑛 ) 𝐶𝑎𝑛 1 − 𝜙 𝑉𝑎𝑛 + 𝜙𝐹𝑎 = 0

-2000

Seed values for 1st iteration: 0.8, 0.2


-3000
Seed values for 2nd iteration: 0.8, 0.5
Seed values for 3rd iteration: 0.65, 0.5
-4000 …. ….
Example Network for UE Assignment

Link performance functions

Link 1 4
𝑉1
𝐶1 = 10 1 + 0.15
2
Link 2 4
𝑉2
O D 𝐶2 = 20 1 + 0.15
4
Link 3
4
𝑉3
Network with three links and one O-D pair 𝐶3 = 25 1 + 0.15
3

Trips between O and D =10


Iteration No. Algorithmic Step Link 1 Link 2 Link 3 Step Size
𝐶1𝑜 =10 𝐶2𝑜 =20 𝐶3𝑜 =25
0 Initialisation
𝑉11 =10 𝑉21 =0 𝑉31 =0
Update 𝐶11 =947 𝐶21 =20 𝐶31 =25
1 Direction 𝐹1 =0 𝐹2 =10 𝐹3 =0  = 0.596
Move 𝑉12 =4.04 𝑉22 =5.96 𝑉32 =0.00
Update 𝐶12 =35 𝐶22 =35 𝐶32 =25
2 Direction 𝐹1 =0 𝐹2 =0 𝐹3 =10  = 0.161
Move 𝑉13 =3.39 𝑉23 =5.00 𝑉33 =1.61
. . . . . .
. . . . . .
. . . . . .
Stochastic User Equilibrium
• The following assumptions made in formulating UE can not be met fully in reality:
• Motorists have full information about the routes
• They make the correct decisions regarding route choice.
• All individuals are identical in their behaviour.
• In reality individual’s perceived travel time is different from the actual travel time.
• Each motorist may perceive a different travel time, over the same link.
• The perceived travel time can be considered as a random variable distributed across the
population of drivers.
• SE definition: Equilibrium will be reached when no traveller believes that his or her travel
time can be improved by unilaterally changing routes. This definition characterizes the
stochastic-user-equilibrium (SUE) condition.
• The stochastic user equilibrium is a generalization of user equilibrium by relaxing the
assumptions made.
Probit Based SUE Assignment

• The perceived path travel time along any given path is normally distributed with a
mean equal to the actual (measured) travel time.
• This path perceived travel time is derived from perceived link travel times.
• In probit based assignment, the link travel times (𝑪𝒂 ) are assumed to be normally
distributed with mean equal to the measured travel time (𝑐𝑎 ) and variance
proportional to the measured travel time:
𝑪𝒂 ~𝑁(𝑐𝑎 , 𝛽𝑐𝑎 ), where 𝛽 is proportionality factor
• The route travel times can be determined as:
𝒊𝒋 𝑎
𝑪𝒓 = ෍ 𝑪𝒂 𝛿𝑖𝑗𝑟
𝑎
MSA Algorithm for SUE
• Step 0: Initialisation. Set n:=1
• Step 1: Sampling. Sample 𝐶𝑎𝑛 from 𝐶𝑎 ~𝑁(𝑐𝑎 , 𝛽𝑐𝑎 ) using Monte Carlo
Simulation for each link
• Step 2: All or nothing assignment. Based on {𝐶𝑎𝑛 }, assign {Tij} to the shortest
path connecting each O-D pair i-j. This step yields the set of link flows {𝐹𝑎𝑛 }.
1 1 𝑛
• Step 3: Flow averaging. Let 𝑉𝑎𝑛 = 1 − 𝑉𝑎𝑛−1 + 𝐹 for all a
𝑛 𝑛 𝑎
• Step 4: Update. Update measured link costs with the current flows {𝑉𝑎𝑛 }
• Step 5: Stopping Test. If flows have not changed in consecutive iterations,
Stop. Otherwise set n = n+1 and go to Step 1.
System Optimum
• Wardrop’s second principle: System optimum
• Under social equilibrium conditions traffic should be arranged in congested
networks in such a way that the average (or total) travel cost is minimised.
• User equilibrium represents user behaviour in choosing routes
• System optimum represents the objective of the planners in
minimising the overall network travel time, thus achieving the social
equilibrium. The drivers are to be persuaded to choose routes in such
away that the average costs are minimised.
• Planners try to achieve this system optimum through congestion
pricing.
Mathematical Programming Framework for System Optimum

• Solution to Wardrop’s System Optimum Problem is obtained by formulating the


following problem:
min 𝑆 𝑇𝑖𝑗𝑟 = ෍ 𝑉𝑎 𝐶𝑎 (𝑉𝑎 )
𝑎

• Subject to
෍ 𝑇𝑖𝑗𝑟 = 𝑇𝑖𝑗
𝑟
and
𝑇𝑖𝑗𝑟 ≥ 0

You might also like