SLIDE 1
The Future of Swiss Railway Dispatching. Deep Learning and Simulation on DGX-1.
Adrian Egli & Erik Nygren
Research and Innovation Platform SBB AG, Switzerland
Adrian Egli & Erik Nygren Research and Innovation Platform SBB - - PowerPoint PPT Presentation
Adrian Egli & Erik Nygren Research and Innovation Platform SBB - - PowerPoint PPT Presentation
The Future of Swiss Railway Dispatching. Deep Learning and Simulation on DGX-1. Adrian Egli & Erik Nygren Research and Innovation Platform SBB AG, Switzerland Swiss Federal Railways. Complex dynamics in the heart of Europe. Basic train
SLIDE 2
SLIDE 3
Swiss Federal Railways. Complex dynamics in the heart of Europe.
SLIDE 4
Basic train dispatching. Reordering of trains.
SLIDE 5
Basic train dispatching. Rerouting of trains.
SLIDE 6
Train runs. A simple chain of dispatching decisions.
SLIDE 7
Interacting trains. The source of railway complexity.
SLIDE 8
Train runs. A path in a decision tree.
SLIDE 9
Most dense mixed train network in the world. Exponential growth of complexity.
1 2 80 4
30 8 Mio. 900
>80
?
~10
80
Mio.
SLIDE 10
Sensitive dynamical system. Finding the needle in the haystack.
SLIDE 11
Increasing future mobility needs. Destabilizing effects of traffic density.
Future Today
SLIDE 12
Maintaining robust traffic flow. Increased man- and computational power.
Future
+ +
SLIDE 13
Maintaining robust traffic flow. Infrastructure enhancements.
Future
+
SLIDE 14
Future projections. Inevitable challenges.
Time Performance
Cost Quality Traffic density
SLIDE 15
Overcoming future challenges. Making the railway network antifragile. Antifragility
SLIDE 16
Antifragility. Improvement through failure.
SLIDE 17
Antifragility. Improvement through failure.
SLIDE 18
How to fail in a safe way. Extending the railway network beyond reality.
Simulation
Validation
Dispatcher
SLIDE 19
Swiss Railway Digital Twin. Infinite possibilities.
SLIDE 20
Reinforcement learning. Mastering complex games.
Game Agent
SLIDE 21
Reinforcement learning. Playing the dispatcher game.
Railway simulation Agent
SLIDE 22
Super human performance. Learning from 65 million years of experience. 65 Mio. years
SLIDE 23
High performance simulations. The power of parallel computations.
python
PyCUDA
SLIDE 24
Digital Twin. Moving beyond the physical boundaries.
SLIDE 25
Digital Twin. Moving beyond the physical boundaries
SLIDE 26
High performance simulations. State of the art.
Reinforcement Learning
13.9 sec.
Business Rules
2.8 sec.
Physics Simulation
0.3 sec.
Swiss railway network
17 sec.
1
31K 15K 13K 800
SLIDE 27
Learning from 65 million years of experience. Time as a limiting factor.
65M years experience 12K years training 17s
1
= x
1
SLIDE 28
Limited time resources. Scaling with innovative ideas.
GPU
Agent Agent Agent
SLIDE 29
Railway simulation. Learning on subregions.
SLIDE 30
Railway simulation. Reinforcement agents view.
SLIDE 31
High performance computing. Parallel training on alternative worlds.
SLIDE 32
Diversity, curiosity, passion and team work. The evolution of a digital twin.
SLIDE 33
Deep learning and simulation. The (r)evolution of the Swiss Federal Railways. Reality Digital Trial & Error
SLIDE 34
Erik Nygren erik.nygren@sbb.ch AI Researcher Adrian Egli adrian.egli@sbb.ch HPC Expert Dirk Abels dirk.abels@sbb.ch Head of Research Lab