Slides for "Towards Interactive Evolution: A Distributed Optimiser for Multi-objective Water Distribution Network Design" presented at HIC2018 - the talk describes the optimisation framework developed as part of the EPSRC-funded HOWS project.
Citation:
D. J. Walker, M. B. Johns, E. C. Keedwell and D. Savic, "Towards Interactive Evolution: A Distributed Optimiser for Multi-objective Water Distribution Network Design", Proceedings of International Conference on Hydroinformatics, 2018.
Abstract:
It is well known that water distribution networks can be optimised by evolutionary algorithms. However, while such optimisation can result in mathematically optimal solutions, the ability of the algorithm to generate novelty can often lead to solutions that are not practical for implementation. This work describes a distributed optimisation platform that will enable the inclusion of a human decision maker in the optimisation process. The architecture of the platform is described, and examples of its execution on benchmark problems is described, using an automated client that interacts with the platform in lieu of a human decision maker.
Towards Interactive Evolution: A Distributed Optimiser for Multi-objective Water Distribution Network Design
1. TOWARD INTERACTIVE
EVOLUTION: A DISTRIBUTED
OPTIMISER FOR MULTI-
OBJECTIVE WATER
DISTRIBUTION NETWORK
DESIGN
D. J. Walker1, M. B. Johns1,
E. C. Keedwell1 and Dragan Savic1,2
1 Centre for Water Systems, University of
Exeter, Exeter, UK
2 KWR Water Cycle Research Institute,
Nieuwegein, The Netherlands
2. Introduction
Evolutionary algorithms are capable of generating highly
mathematically optimal solutions that are often
unsuitable for implementation in the real world
• Missing objectives
• Missing constraints
• Changing objectives/priorities over time
Solution: involve an engineer in the
optimisation process – interactive
evolutionary algorithms
3. Interactive EAs
Interactive EAs generate solutions by combining human
expertise with evolutionary processes
•Selection: user evaluates solutions (no fitness functions)
•Mutation: user modifies solutions (no mutation operators)
picbreeder.org
Generation 1 Generation 2
5. Server-side Optimiser
The server has a modular design so that different
optimisers and solvers can be used
•Currently uses a single-point optimiser and EPANET
•Client applications currently supports a single solution
•Keeps an archive of solutions – estimated Pareto front
6. Message Passing - JSON
JavaScript Object Notation is a standard format for data
interchange over the web.
• Network information
• Problem-specific information (e.g., available diameters)
• Start a new session
• Initiate optimisation
• Look-ahead
{"diameters": [{"id": "1", "diameter": 144.0}, {"id": "2",
"diameter": 156.0}, {"id": "3”, "diameter": 156.0}, {"id":
"4", "diameter": 108.0}, {"id": "5", "diameter": 132.0},
{"id": "6", "diameter": 48.0}, {"id": "7", "diameter": 84.0},
7. Testing the Server
The server has been tested with a simple client that sends
JSON messages to optimise a benchmark WDN
1. Minimise network cost
2. Minimise head deficit
Look-ahead
• Client allows user to view all
possible interventions
• Test program automates this
• Compares with and without look-
ahead
8. New York Tunnels
Combined solution sets for 30 repeats with look-ahead
and without look-ahead - Look-ahead has produced a
tighter distribution of solutions
9. Hanoi
Combined solution sets for 30 repeats with look-ahead
and without look-ahead - Look-ahead is still slightly ahead
(though by a less clear margin than for NYT)
12. Conclusions
• The HOWS project: putting the engineer back into the
loop.
• Client-server methodology: modular approach allows
for substitution of different optimisers, solvers and
clients.
• Light-weight message passing: simple JSON
messages prevent the application from slowing down
while messages are exchanged.
• Multiple clients maximise usability: 3D visualisation,
visualisation in the browser, virtual reality.
13. Future Work
• More complex problems: we are working on including
more complex networks and their components (e.g.,
pumps).
• Modelling user behaviour: a problem with interactive
evolutionary algorithms is user fatigue – we are
investigating model-based approaches to dealing with
this (paper accepted to CCWI 2018).
• Expanding the user interface: providing the user with
more control of the network, and expanding the
application to work in VR.