Fix LMMAES #56
Fix LMMAES #56
Conversation
| @@ -158,7 +138,7 @@ def ask_strategy( | |||
| state.c_d, | |||
| state.gen_counter, | |||
| ) | |||
| return x, state | |||
| return x, state.replace(z=z) | |||
There was a problem hiding this comment.
Maybe I am missing it, but where is the z needed? (popsize, num_dims) will require quite a lot of memory for many applications. Can we reconstruct this from the sampled x candidates instead?
There was a problem hiding this comment.
Yes, this modification is the largest difference between original evosax's code and numpy code for memory size.
We have to use z to update p_sigma,M(pls see the update functions), but in order to get z from solution vectors we have to calculate inverse matrices and that may causes both computing errors and costs. I think the procedure for getting z from solutions will be like this;
iteratively calculate inverse of (1-c_d[j])*I+c_d[j]*jnp.outer(M[j,:],M[j,:]) like sample function and select j by checking generation.
Are there any linear algebra tricks to keep away from calculation of inverse matrices? I can't came up with that.
There was a problem hiding this comment.
Mhmm, I am not sure -- but I guess this is a bit of a problem. I guess it will be hard to scale to neuroevolution even though it is the limited memory version :( I will give this more of a thought.
There was a problem hiding this comment.
Hello. If possible, could you please approve the workflow and run the test before devising the implementation? I am wondering if there are any implementation mistakes at this time.
add update best fitness and best member.
updating best_fitness and best_menber are done in Strategy's tell function so we needn't it.
updating best_fitness and best_menber are done in Strategy's tell function so we needn't it.
fix type check typo
Codecov Report
❗ Your organization is not using the GitHub App Integration. As a result you may experience degraded service beginning May 15th. Please install the Github App Integration for your organization. Read more. Additional details and impacted files@@ Coverage Diff @@
## main #56 +/- ##
==========================================
- Coverage 88.86% 88.85% -0.01%
==========================================
Files 76 76
Lines 4796 4793 -3
==========================================
- Hits 4262 4259 -3
Misses 534 534
|
This pull request is for this issue.
Compared with implementation in numpy, the original LM_MA_ES in evosax seems to be incorrect wrt default parameters settings and parameters update expressions so I write a code corresponds to numpy implementation.
I used black linter and tested by evosax/tests/test_strategy_api.py. This is my first contribution, so please tell me if there are any problems.