Lab2 Review

[saccuz]

General Thoughts

Well done in providing more than one way to mutate and to crossover.

Disagreement

• Considering you had already different ways to perform mutation and crossover, you could have added ways to tweak at runtime the probability of using one instead of another in order to better handle exploitation or exploration w.r.t the situation.
• Applying mutation on the result of a mutation could lead to wasting some usefull part of genome.
• Maybe the loseweight mutation is a bit not context agnostic?

Conclusion

Great work, you take advantage of the python language to randomize the use of different kind of mutation and crossover. Results you found are absolutely good, also considering the low amount of time needed to obtain them.

[LeonardoTredese]

I agree that the loose weight mutation is context biased, I did that on purpose. I wouldn't always be useful.

[LeonardoTredese]

You are right, I could use an adaptive mutation rate between the two mutations and crossovers. I would also try to avoid the double mutation and see if the performance improves.

[LeonardoTredese]

I ran the code in different settings to check the behavior of applying multiple mutations in the same epoch. For each problem size I repeat the search three times

With double mutation

	N=5	N=10	N=20	N=100	N=500	N=1000
Mean weight	5	10	24	188	1518.33	3662.33

Performing population mutation after offspring generation

	N=5	N=10	N=20	N=100	N=500	N=1000
Mean weight	5	10	24	210	1442	3472.33

Here I was surprised that for N = 500 the result was exactly 1442 all three times

Performing population mutation after generating offspring, No mutation when creating offspring.

	N=5	N=10	N=20	N=100	N=500	N=1000
Mean weight	5	10	24	200.33	1665.33	3962.33

The results for N < 500 are all pretty similar, so I will consider only larger results. Performing the population mutation independently from offspring seems to improve the solution. Still it appears that applying a slight mutation after the crossover is still useful.

[LeonardoTredese]

I tried to reduce the number of crossovers while in further generations at using the rate $\frac{\alpha}{\sqrt[]{n}}$ where n is the generation number.

	alpha	N=5	N=10	N=20	N=100	N=500	N=1000
Mean weight	10	5	10	25	248	1859.33	4265.66
Mean weight	5	5	10.33	24	201.33	1634	3718.33
Mean weight	2	5	10	24	201.66	1684.33	3860.66
Mean weight	1	5	10	25.33	213	1595.33	3877.33

Apparently reducing the crossovers worsens the best found solution, therefore I will not keep it in the final implementation. Probably because the algorithm will explore less search space.

Comment on "Performing population mutation after offspring generation"

At the light of what was explained at lesson, in the previous comment I noticed that performing a mutation with a certain probability after the crossover improves the final solution. A possible explanation is that the offspring will be different from the parents even if they are similar or identical therefore improving diversity and reducing the quick convergence rate caused by the loose weight mutation.