Correcting a parameter, adding parallel processing, and experimenting with parameters to try and run similarly but more quickly. Total time used: 11h 21m (28.4%)
Would’ve been handy if all parameters could have been mentioned in one place.
| Parameter | Code | Location in paper | Location in code |
|---|---|---|---|
| Length of simulation: 1 hour | ✅ self.maxTime = 60 |
4.3 Processing | PODSimulation() __init__ in PodSimulation.py |
| At screening station, 1% go to med eval and 99% to dispensing. At med eval station, 99% got to dispensing and 1% exit POD | ❔ Can’t find | 4.1.1 Splits | - |
| Number of forms per designee: 1 31.8%, 2 26.7%, 3 16.8%, 4 12.6%, 5 6.8%, 6 5.6% | ❔ throughput x 3.2 - so appears to be averaged to 3.2? | Table 1 | plotting_ staff_ results.r |
| Service time line manager: triangular, minimum = 0.029, maximum = 0.039, mode = 0.044 | ❌ time = random. triangular( low=5/60.0, high=92/60.0, mode=23/60.0). Low 5/60 = 0.0833. High 92/60 = 1.533. Mode 23/60 = 0.3833. There is one commented out which has… Low 1.77/60 = 0.0295. Higher 2.66/60 = 0.044. Mode 2.38/60 = 0.0397. This would match the article, except maximum and mode the other way round. |
Table 2 | visit_greeter() in Customer.py |
| Service time screening: weibull, shape = 2.29, scale = 0.142 | ✅ time = random. weibullvariate( alpha=0.142, beta=2.29 ) |
Table 2 | visit_screener() in Customer.py |
| Service time dispensing: weibull, shape = 1, scale = 0.311 | ✅ time = random. weibullvariate( alpha=0.311, beta=1 ) |
Table 2 | visit_dispenser() in Customer.py |
| Service time medical evaluation: lognormal, logarithmic mean = 1.024, logarithmic stdev = 0.788 | ✅ time = random. lognormvariate( mu=1.024, sigma=0.788 ) |
Table 2 | visit_medic() in Customer.py |
| Arrival rate 100 designess per minute per POD (following a Poisson distribution) | ❌ self.meanTBA = 1/200.0 #1/float(115) #mean time between arrivals, minutes btw entities. This would mean 200 arrivals per minute, rather than 100. |
4.1.4 Arrival rate | PODSimulation() __init__ in PodSimulation.py |
| Three simulation runs | 🟡 This wasn’t the case when I first started, but have already been changing this | 4.3 Processing | main.py |
| Number of staff members per station - mentions examples of where “each station could have up to thirty staff members” or “for example 60”. We know it cannot be 30, but could reasonably assume to be 60 | 🟡 This wasn’t the case when I first started, but I have already noticed and addressed, and fixed to 60. | 3 Problem and 4.3 Processing | Staff Allocation Problem() in Staff Allocation Problem.py |
| Default crossover rate 1.0 and n=1 | ✅ ea.variator = [variators. n_point_crossover] with variators imported from inspyred.ec, which can see from docs the default crossover rate is 1 and default number of crossover points is 1 |
5 Experimental results | nsga2.py |
| Experiment 1: tri-objective model, population 100, generations 50, pre-screened scenarios 10%, 20%, 30%… 90% | ✅ As in the input files like 10-prescreened.txt |
4.1.1 Splits and Figure 5 | As left |
| Experiment 2: bi-objective model, population 50, generations 25, pre-screened scenarios 10%, 20%, 30%… 90% | TBC | 4.1.1 Splits and Figure 7 | - |
| Experiment 3: tri-objective model, pre-screened percentage ??, (a) 100 pop 50 gen (b) 200 pop 100 gen (c) 50 pop 25 gen | TBC Note: Where I am unsure of pre-screened percentage here, I presume it might be default from code which, if parameterReader == None, then self. preScreened Percentage = 0.1 |
5.3 Experiment 3 and Figure 8 | -PODSimulation() __init__ in PodSimulation.py |
| Experiment 4: tri-objective model, maximum line managers 1, 2 or 3 | TBC | 5.4 Experiment 4 and Figure 9 | - |
| Experiment 5: 6 dispensing, 6 screening, 4 line manager, one medical evaluator, number of replications 1-7 | TBC | 5.5 Experiment 5 and Figure 10 | - |
Reflections on discrepancies (❌): Big difference in the line manager service times - this might explain it! Will start with this, but could then try some of the others? Wary of trying all at once, as always hard to know what might be right - the code or the article. Will run 100pop 5gen 1 run.
This took 210 minutes to run (3 hours and a half). The Y axis of Figure 5 is now much higher - in fact, too high!
Figure 5:
Figure 6, filtered to just the throughouput in the range of those plot in the article:
I tried switching to 10pop 1 gen 1run, just to confirm if it gives results with similar range, as if so, that’s much quicker for trying out different changes, but these looked rather different!
Run time: 6 minutes
These looked quite different though, so will need to run with a bit more…
Figure 5:
Figure 6:
I tried switching the loop in Experiment1.py into a parallel loop, to speed up the run time.
First, I just changed it to a function with a loop.
I ran this with the same parameters as above (10 pop 1 gen 1 run) to confirm the results didn’t change at all due to the parallel processing and, indeed, they remained the same, so this is successfully implemented. The only change was - as expected - to the times files, which is now a single file.
Run time: 6 minutes
I then adjusted the loop to use parallel processing (with multiprocessing Pool). This required the old syntax (i.e. not the with() statement).
Run time: 55 seconds
The pre-existing seed control and way the code was structured made it really easy to implement this.
I then tried running with 50 pop 25 gen 1 run, as they used that for Figure 7 (though 3 runs) so it should be similar.
Run time: 57 minutes
I realised then that that had just as much impact on the y axis as anything else! Hence, I figured best plan of action would be to run as is with 100 pop 50 gen 1 run, and see how that looks. Then, if that doesn’t look right, try tweaking with parameters identified in table above.
import sys
sys.path.append('../')
from timings import calculate_times
# Minutes used prior to today
used_to_date = 606
# Times from today
times = [
('09.14', '09.47'),
('13.15', '13.25'),
('13.26', '13.35'),
('13.44', '14.03'),
('15.17', '15.21')]
calculate_times(used_to_date, times)Time spent today: 75m, or 1h 15m
Total used to date: 681m, or 11h 21m
Time remaining: 1719m, or 28h 39m
Used 28.4% of 40 hours max