[ProgSoc] Predict Sydney traffic?

John Elliot jj5 at jj5.net
Sun Dec 5 02:15:29 EST 2010


On 3/12/2010 12:17 AM, marauder wrote:
> More generally we say that time series are made up of four kinds of
> thing stacked on top of each other: trends, cycles, seasonality,
> and irregular movements.

I had a guess that M4 traffic speed over a given route could be made up 
of the following:

  * most recent speeds (trend)
  * daily cycles (cycle)
  * weekly cycles (cycle)
  * least squares estimate based on last n samples (trend)

I created a formula which factors these components into a single 
estimate, and created a genetic algorithm which modifies the factors 
that control the significance (and other features) of each component. I 
ran my genetic algorithm for a night and got down to a model that is 
wrong by 2 seconds on average, which I suspect is a fairly good estimate.

My code is here [1]. I think I'll make a submission for the competition, 
and if I win I'll donate $1337 to ProgSoc.

One problem with my code is that when I stumble upon a good set of 
parameters for my model I have trouble recording them, because the model 
is sensitive to minor variations in double precision floating point 
values, and recording those values as a string loses some precision. I 
think I'll have to put in a serialisation system that can record the 
precise values.

[1] http://www.progsoc.org/~jselliot/rta-m4/rta-m4-2010-12-05-013900.zip










More information about the Progsoc mailing list