Added simple_ga.py algo file

[MaximilienLeClei]

No description provided.

[lerrytang]

By setting best_params, we expect the algorithm can continue training from that point, it seems self._best_params is not used in this way.

[MaximilienLeClei]

Should work as intended now.

[MaximilienLeClei]

Actually I'm a little confused about the format of best_params. Is it the parameters of the elite agent or the whole batch of parameters from the entire population?

[lerrytang]

It should be the top elite agent's parameter.
Users call NEAlgorithm.best_params to save/test the model.

[MaximilienLeClei]

Got it and fixed the setter!

[lerrytang]

Hi, thanks for the PR, I'm testing its performance on the tasks.
At the same time, can you take a look at my review comment and make changes accordingly?
Please see our implementation as an example.

[lerrytang]

Test results

	Benchmarks	Parameters	Results
MNIST	90.0 (max_iter=5000)	sigma=0.001	90.8
CartPole (easy)	900 (max_iter=2000)	default	925
CartPole (hard)	600 (max_iter=2000)	default	616
Waterworld	6 (max_iter=2000)	default	6.39
Waterworld (MA)	2 (max_iter=5000)	default	1.19
Brax Ant	3000 (max_iter=1000)	truncation_divisor= 4	2336

Notes

1. This table shows the implementation's performance so that users can choose the algorithms for their experiments.
2. We are aware that some algorithms have limitations (e.g., unable to train large policy network), the benchmarks are therefore not hard requirements. However, we refuse to merge if some scores are significantly lower.
3. After we release the test scripts, the PR submitter will be responsible for producing this table. We can help fill in some entries if the submitter cannot run the experiments due to hardware limitations.