Monte-Carlo simulation of a bootstrap particle filter
This weeks homework will be a competition. The goal is to optimize the code detailed below, the winner will be the one who has provided the fastest running program (as determined by me running on my machine billman) that is
- Submitted in time.
- Equivalent to the original code, in the sense that it can reproduce the estimation performance demonstrated below, without fundamentally changing the particle filter algorithm or the number of particles/time-steps.
- I provide the baseline code
- My code provides a decent particle filter implementation
- The code is bad from a julia-performance point of view
- Your job is to optimize it
- Optimized code has to be equivalent (do not implement different algorithm)
- If you get stuck, find me in my office and I'll help you out.
The output of the program will look something like this
N = 5
N = 30
N = 100
N = 300
N = 1000
N = 10000
T = 20
N = 5
N = 30
N = 100
N = 300
N = 1000
N = 10000
T = 200
N = 5
N = 30
N = 100
N = 300
N = 1000
N = 10000
T = 2000
Propagated 359755800 particles
168.876344 seconds (926.23 M allocations: 115.608 GiB, 10.36% gc time)
The code took approximately 170 seconds to run on my machine and produced the following result
It is a good idea to continuously test your code and make sure you maintain a similar estimation performance as you modify the code.
My optimized version runs in about
8.525856 seconds (1.02 M allocations: 172.929 MiB, 0.53% gc time)
for an approximately 20-fold reduction in running time (and a 700-fold reduction in allocated memory). This will give you an idea of at least how far it is possible to push it.
-Fredrik