Hyperparameter search algorithms

[hill-a]

Long time without a pull request, this fixes that.

• added Hyperband
• added Hyperopt
• made all the RL algorithms compatible with the hyper parameter search
• optimized cma-es hyperparameters
• optimized PPO2 lstm hyperparameters
• validated PPO2 hyperparameters for feedforward policy with kuka and mobile robot
• added interface class for gym racing car environment
• added PPO2 for the dataset generator
• fixed a few multiprocess issues with environments
• added a table and animations for all the environments in the README
• upgraded tests, and can now call a fast test run with ./run_tests.sh --fast (only ground truth with all the RL algorithms and all the environment tested on ground truth), and call all the tests with ./run_tests.sh --all (--all is with hyperparam search + enjoy script testing). Also ./run_tests.sh --help will print out all the possible flags that pytest can use.

TODO:

• [x] find good hyperparameter for PPO2 LSTM policies
• [ ] find good hyperparameter for A2C, ACER, DDPG, DeepQ, SAC and ARS

[araffin]

Cool =), you will have to add sac soon ;)

[hill-a]

tests: 17 passed in 5386.25 seconds for 012f2d6, with 85% coverage

[araffin]

@hill-a do you plan also to add something on the tests to cover the hyperparameter optimizers ?

[hill-a]

Dont merge the pull request yet, the tests show an issue with hyperopt.

[kalifou]

As discussed @hill-a, for the table about "Available Environments" in the Readme, we should mention also that the GT can be relative position from Robot to Target (in footnote for instance)

[hill-a]

Will run ./run_tests.sh --all before tomorrow, to validate the tests. Ready for review again.

[hill-a]

tests --all: 83 passed in 9707.51 seconds for e8783ee, with 86% coverage