Create custom network

MyGym toolkit is available in such form, that enables you to start with your own training very easily and quickly. You can select one of the Reinforcement learning algorithms implemented by Stable baselines 1 or try to design your own. This tutorial shows how to train a custom network in myGym with your own implementation of a learning algortihm. We recommend you to get familiar with the structure of Stable baselines implementations first.

To create your custom algorithm we recommend you to define it as a child of an existing Stable baselines algorithm class. To create your custom policy class, follow these instructions.

Choose a keyword for your custom algorithm and add new value to the dictionary of implemented combinations in train.py:

implemented_combos = {"myalgo": {*framework*: [MyAlgoClass, (MyPolicyClass, env), {*training kwargs*}]},
                        ...}

To start training with your custom network, set your new algorithm keyword as the algo parameter in the config file:

"algo":"myalgo",

or in the command line:

python train.py --algo=myalgo

The function train in the train.py script will read the value from the implemented_combos dictionary:

model_args = implemented_combos[arg_dict["algo"]][arg_dict["train_framework"]][1]
model_kwargs = implemented_combos[arg_dict["algo"]][arg_dict["train_framework"]][2]

create a model, an instance of MyAlgoClass:

model = implemented_combos[arg_dict["algo"]][arg_dict["train_framework"]][0](*model_args, **model_kwargs)

and start with training:

model.learn(total_timesteps=arg_dict["steps"], callback=callbacks_list)

When finished, trained model is saved to the lodgir:

model.save(os.path.join(model_logdir, model_name))

1

Hill, Ashley and Raffin, Antonin and Ernestus, Maximilian and Gleave, Adam and Kanervisto, Anssi and Traore, Rene and Dhariwal, Prafulla and Hesse, Christopher and Klimov, Oleg and Nichol, Alex and Plappert, Matthias and Radford, Alec and Schulman, John and Sidor, Szymon and Wu, Yuhuai (2018). Stable Baselines. Github repository.