Approximating Traffic Simulation using Neural Networks and its Application in Traffic Optimization Appendix 1: Results of experiments

Paweł Gora University of Warsaw [email protected]

Karol Kurach University of Warsaw [email protected]

Parameters Size of the training set:

93626 (80% of all data)

Size of the test set:

23406 (20% of all data)

117033 traffic signal settings evaluated Training and test sets generated using in a microscopic model, 10 minutes of Traffic Simulation Framework software simulation, 30000 cars initially, 20 new cars each second Batch size:

100

Cross-validation:

5-fold

Results Learning rate = 0.001

[100, 100, 100] 3 layers 100 neurons / layer

[100, 100] 2 layers 100 neurons / layer

[200, 200] 2 layers 200 neurons layer

[200] 1 layer 200 neurons

Avg relative error

2.03%

2.39%

1.75%

4.44%

Max relative error

10.08%

13.61%

8.89%

22.78%

Results Learning rate = 0.01

[100, 100, 100] 3 layers 100 neurons / layer

[100, 100] 2 layers 100 neurons / layer

[200, 200] 2 layers 200 neurons layer

[200] 1 layer 200 neurons

Avg relative error

1.56%

2.28%

2.11%

3.98%

Max relative error

9.18%

11.72%

10.8%

20.06%

Results Learning rate = 0.1

[100, 100, 100] 3 layers 100 neurons / layer

[100, 100] 2 layers 100 neurons / layer

[200, 200] 2 layers 200 neurons layer

[200] 1 layer 200 neurons

Avg relative error

1.74%

2.62%

2.01%

3.76%

Max relative error

8.47%

13.31%

11.31%

18.64%