PTSP Competition Session GECCO 2005 by sdfwerte


									PTSP Competition Session
     GECCO 2005
    Session Chair: Graham Kendall
  Competition Organiser: Simon Lucas
    Physical Travelling Salesman
Visit all cities in minimum
Each time step, choice of
5 force vectors
Solutions are strings over
Secondary criteria:
   Minimise non-zero forces
   Given solutions of equal
    quality, earliest submission
Number of entrants:
   10 (by unique email addresses)
   Plus many anonymous (not necessarily
    different) entrants
Number of entries: 68
No limit on the number of entries per
        Top Ten Entries
Jiaqiao Hu = umd; Jose Martin = IAI CSIC
Jose Martin
Jiaqiao Hu
     Rok Sibanc (2nd Place)
Best solution: 652, 652, Wed Jun 22 12:49:15
                   Rok’s Method
    (Summarised by Simon Lucas with apologies for any errors)

Stage 1:
   Optimise route (permutation of cities)
   With a novel criteria
   Includes angles between consecutive straight-line
   As well as distance
   And a weighting parameter to balance contribution of
Optimised with a population-based, mutation
only EA
Tournament Selection
Rok’s Force Vector Optimiser
Stage 2: for a given route
   Find best set of forces
Interesting: uses only symbols 1-4
   (the unit vectors)
Never uses a zero-force step
Incremental approach:
   Optimise a solution that visits just the first city,
   Then the first two, then three etc…
Mutation only EA
       Rok’s Fitness Function
The fitness function:
   • iSeen is the number of visited cities
   • iUsed is the number of used vectors
   • m_iGenomeMaxLenght is the maximum number of
    vectors for a genome
   • dNearest is square of the distance of last position
    before last visited city
   • dLastDist is square of the distance of last position to
    the next unvisited city
  Rok’s Variation Operators
1-Symbol Substitution
1-Symbol Deletion (and insertion?)
Replace a subsequence of length n, with n
copies of the same symbol
Deletion of a subsequence
Probability of these mutations such that
small changes more likely than large ones
     Martin Byrod – 1st Place
Winning entry: 648, 636, Wed Jun 22 12:41:11
              Martin’s Method
  (Summarised by Simon Lucas with apologies for any errors)

Two stage:
Use a standard EA to optimise the route,
given standard TSP cost function
Then plug in the real cost function (which
may alter the TSP-estimated route)
The real cost function is determined by an
EA that optimises the solution string for a
given route
Population-based EA, mutation only
    Martin’s Fitness Function
Fitness function:
N: Number of cities visited
E: distance to next city
T: total time taken (length of
Optimised using a greedy
sliding window approach
0 to 100, then 10 to 110 etc.
Force Vector EA initialised
with routes derived from PD
     Key Idea: Swap Mutation
Use both Swap (right) and bit-flip (left) mutation
Swap makes much smaller changes
But bit-flip needed to make all strings reachable
Sample swap: 1234 -> 1324
       Bob MacCallum (6th) : GP
    (Summarised by Simon Lucas with apologies for any errors)

Used PerlGP to evolve a controller
Different approach to the other methods
The controller takes as input:
   Current state (position, velocity)
   Locations of cities that are yet to be visited
Outputs a continuous force vector at each
   Which is then quantised to form next solution
Bob’s Best Route
PTSP: Interesting challenge that stimulated a good deal
of interest
Many different approaches possible:
   Some work much better than others
   Simple naïve methods perform poorly
Web-based continuous league:
   Interesting to observe
   Psychological aspects also
Competitors showed fantastic ingenuity!
Thanks to all participants for making it a worthwhile
Might run a future contests where algorithms instead of
solutions are submitted…

To top