EXPERIMENTING WITH INTELLIGENT TRAFFIC SIGNAL CONTROL

1
EXPERIMENTING WITH INTELLIGENT TRAFFIC SIGNAL
CONTROL

• Alenka Malej MSc
• Andrej Brodnik PhD
• University of Primorska
• PINT
• may 2007

2
Motivation

• optimization of traffic signal control for a
  network of intersections
• cost effectiveness of improving traffic flow
  through the network
• flexible, easy to understand, intuitive,
  adaptive, operates with uncertain data
• real-time
• proactivity
• on-line learning

3
Categories of control logic

• pretimed signal phases, offsets, cycle length
  determined off-line
• vehicle-actuated extensions of signal phases in
  response to vehicle actuations
• traffic responsive control vehicle actuation,
  prediction, pattern matching

4
Responsive traffic signal control

• traditionally commom cycle length
• distributed control allows different cycle
  lengths
• a matter of adaptability incremental adjustments
  to signal parameters

5
Intelligent vs. Adaptive

• What is adaptive?
• traffic responsive
• real-time,
• proactivity acting in advance, to deal with
  expected traffic conditions(mathematical models
  or AI)
• Intelligent improve over time

6
Artificial intelligence

• Fuzzy logic knowledge representation
• reinforcement learning learning algorithm
• Fuzzy Q-learning (FQL)
• distributed control multi-agent systems

7
Fuzzy logic
8
Reinforcement learning

• deals with interactions, consequences actions
  have influence on the environment (-gthistoric
  data is not enough)-gt RL suitable for traffic
  signal control
• uses feedback - reinforcement function for
  rewarding actions in given states
• state values longterm optimization

9
Past studies

• Bingham isolated control (2001)
• Bogenberger ramp-metering (2003)
• Networks of intersections
• Chiu Chand (1993) distributed fuzzy control
  system, cycle, split, offset
• Nakamiti, Freitas (2002) case-based mehanizem, 1
  s extensions
• Choy et al. (2003) neuro-fuzzy, genetic-fuzzy
  algorithm, cycle, split, offset

10
Our model

• based on Chiu and Chand, modified and extendex to
  FQL
• object of control cycle, split, offset
• level of control distributed
• detectors position stop bars
• type of control hybrid system (fuzzy Q-learning)

11
(No Transcript)
12
If occupancy is high and occ_change is
positive, then cycle_change is positive. If
occupancy is Veryhigh and occ_change is
negative, then cycle_change is negative.
13
Split rule example If occupancy is not(high) and
occ_diff is positive, then split_change is
positive.
14
Q-matrix example
consequences (13)
rules (5)
Exploration Exploitation Policy -gt which
consequences (and therefore action) will be
chosen in a given state
15
Reward
target occupancy 50 /- 10
16
Simulations

• choice of simulation program !!!CORSIM
• external signal control aplicationRun-time
  extension (RTE DLL)

17
Results of past studies

• Choy Singapur, morning peak hour
• 15 smaller total delay
• compared to SCATS

18
Simulation results Koper
24 smaller total delays
all simulation results are compared to simulation
results with pretimed control optimized with
TRANSYT (genetic algorithms)
19
18 smaller total delays
20
3 x 3
55 smaller total delays
21
3 x 3, less detection
17 smaller total delays with only 4
intersections with detectors
22
3 x 3 ... learning ...
22, 36 smaller total delays
23
Ljubljana
16 smaller total delays (less detectors and
other limitations)
24
Results
15 years, d.r.3.99, 1h 1,6 EUR
GDP/population hours
B/C gt 1 also in one year! minimum for B/C gt
1 2,5 smaller delays OR 0,8 fuel
savings sensitive on start of benefits, one
year delay -gt 9 smaller B/C, but BgtC allways
25
B/C depending on number of intersections
26
Questions?