donghwan-shin
commented
2 years ago What I did
So, I found three different issues, addressed in the following commits:
4af1f3b33ed3875b7d504260e12c5fc10aaf85dd
This is to fix a bug in calculating the safety_req_value value of each complete solution. Previously, the value was zero for all complete solutions since the map() function you used did not properly save the results in the corresponding field.
698f59565f10d749a0d1b7b5aa8878c310c53727
This is to reduce the "radius" (i.e., max_dist) of a cs_region; since 0.5 was too large, the search does not make fine-grained movements and is easily stuck in some regions.
eebc72f9ff57df4a77e39e0e291849e5d2a6352f
This is to make the MTQ problem "solvable". Previously, use used h2= 150 but s_2 = 1.0/32.0, meaning that it is very challenging to find the peak of h2 on a very small surface (1/32). So, I simply swapped h1 and h2 to make the problem easier to solve.
New result and its interpretation
This is the full log: 20220609_160328_CCEA_MTQ.log
As you can see, we have reached the fitness score of 0.9735897570338863.
Using the final solution x, y = [0.5853844653731718], [0.894080760273642], the function compute_safety_req_value() returns -21.786309582260785, which is quite close to 0.
In fact, on different runs, I found x, y = [0.6182038195955862], [0.8721837981696042], resulting the return value of compute_safety_req_value() only 1.5513294434849598.
Since the positive return value of compute_safety_req_value() indicates "safe" and the negative one indicates "unsafe", the return value of zero indicates the "boundary". Therefore, we can say that the algorithm successfully found a complete solution (x, y) that lies near the boundary.
To conclude, the sanity check is passed, but it turned out that max_dist should be carefully tuned for a given problem
SepShr
The MTQ benchmark problem will act as the first step. To run the benchmark problem simply run the following command:
python3 run_iccea_mtq.pyTo modify the search parameters you can change them in
benchmark/mtq/search_config.py.To modify the boundary, you can change the value at the end of the
joint_fitness_mtq()function located inbenchmark/mtq/problem.py. Kindly note that the joint fitness function does not return the actual joint fitness value in the updated version of the ICCEA, rather it returns the value of thesafety_req_valueattribute of the complete solution.