Master AIC - RL courses - Evolutionary RL

Organized by herilalaina - Current server time: Jan. 21, 2021, 5:13 a.m. UTC

First phase

Jan. 13, 2019, midnight UTC


Competition Ends
Feb. 4, 2019, 11 p.m. UTC

Master AIC - RL courses - Evolutionary RL

Brought to you by University of Paris-Saclay

Evolutionary RL, Practical session 4

In this practical, you are asked to put what you just learnt
about direct policy search. You are provided with the `` file. Use `python -h` to check how you are supposed to use this file.

In this project, you are asked to solve the classic Mountain Car ( For more details about action and observation space, please refer to the OpenAI
documentation here:

How do I proceed ?

As previous pratice, the baseline (RandomAgent) is provided as a starting kit.   It
contains several files, two of them are of interest for you: and is the file in which you will write the code of your agent,  using
the RandomAgent class as a template.  Don’t forget to read the documentation
it contains.  As usual you can have the code of your several
agents in the same file, and use the final line `Agent = MyAgent` to choose which agent you want to run.

You are expected to implement direct policy search algorithm using evolutionary computation (CMA-ES).
The running of your agent follows a general procedure:
1. Train your agent using CMA-ES in `train` method
2. Test and score your final agent using `test` method

We are in the setting of model free approach (Pg 10.

In order to efficienlty train your agent, you must:
* Define your search space (policy space in which your are willing to search for)
* Define your objective function: to assess a policy (Episode-based or step based)
* Optimize the objective using cma-es: use Be aware that we setup a time budget of 2hr in Codalab. You can always start doing local train then initialize your agent with good values for Codalab submission.

Challenge your friends
You are required to have a codalab account (
Please use an username in the format of *firstname.lastname* when creating your account.
You can enter to the competition here

For submission, you need to zip `` and `metadata` files then submit the zipped file to codalab.
`` as an example of submission.

If you want to reproduce your local score on Codalab, please use the docker image ( and do not change the seed.
Then run `python --ngames 1000 --niter 1500 --batch 10`

For further questions, please use the codalab forum.

* 3/4: Agent implementation
* 1/4: rank on the competition

Deadline: 28/01/2019

Only the last submission is considered for grading (agent implementation) and ranking (challenge).

Master AIC - RL courses - Evolutionary RL: Evaluation

The final performance of your agent will be evaluated on a 2000 random testbed. The average reward is considered as metric.

You can assess your algorithm locally with the following command:

python --ngames 500 --niter 200 --batch 10


Start: Jan. 13, 2019, midnight

Description: Development phase: create models and submit them or directly submit results on validation and/or test data; feed-back are provided on the validation set only.


Start: Feb. 3, 2019, 11 p.m.

Description: Final phase: submissions from the previous phase are automatically cloned and used to compute the final score. The results on the test set will be revealed when the organizers make them available.

Competition Ends

Feb. 4, 2019, 11 p.m.

You must be logged in to participate in competitions.

Sign In