rewards

A low code sdk for crearing custom environments and deep RL agents.

Installation

[linux]

Installing rewards is easy in linux. First clone the repository by running

git clone https://github.com/rewards/rewards.git

One cloned go to the repository and make sure make is installed. If not installed just run:

sudo apt install cmake 

Once done, now create a new virtual environment and install dependencies. You can achieve this by running the following:

make virtualenv
make install 

This should install all the dependencies and our sdk rewards:v1.0.0.

[windows]

For installation in windows, it's also simple. All you have to do is just clone the repository same as before. Then create a new virtual environment.

virtualenv .venv

Load the virtual environment

.\venv\Scripts\Activate

Now go to the repository and install all the dependencies and the rewardss package.

pip install -r requirements.txt
python setup.py install

Getting started

rewards is mainly made for two important reasons.

• First we want to make learning reinforcement learning easy, by introducing this low code framework. So that folks do not need to spend more time in making environments or other stuff. All they can focus is on creating different agents, models and expeiment with them.

• We want to make it as interactive and begginer friendly as possible. So we are also introducing rewards-platform which where we gamified the experience of learning RL.

• If playing games can be fun and competitive then why not RL? Hence with rewards-platform and rewards you can host and join ongoing competition and learn RL with your friends.

NOTE: Our coming enterprise version is mainly focussed to build the same but for RL/Robotics based companies where we want to ensure that their focus lies more on the research rather creating environments and other configurations.

Take a look on how to get started with a sample experiment

Currently this version of rewards only supports a single game and environment. That is car-race. We will be adding support for more environments (including gym, unity, and custom environments) very soon.

So let's go ahead and see how to get started with a sample experiment.

from rewards import workflow

configs = workflow.WorkFlowConfigurations(
    EXPERIMENT_NAME="Exp 3", 
    MODE="training", 
    LAYER_CONFIG=[[5, 64], [64, 3]]
)


flow = workflow.RLWorkFlow(configs)
flow.run_episodes()

First you call our sdk's workflow module. The workflow module helps us to

• Create Environments and configure environments
• Create models and configure them
• Run the whole experiment and log all the results

All at one place. We first get started with writing our own configuration using

configs = workflow.WorkFlowConfigurations(
    EXPERIMENT_NAME="Exp 3", 
    MODE="training", 
    LAYER_CONFIG=[[5, 64], [64, 3]]
)

Here is the table of configuration and what they means

So above is a quick overview of how to use different reward configurations. Now once configuration part is done, load those configuration to RLWorkFlow() and run for a single episodes.

After this you are ready to run the above code:

from rewards import workflow

configs = workflow.WorkFlowConfigurations(
    EXPERIMENT_NAME="Exp 3", 
    MODE="training", 
    LAYER_CONFIG=[[5, 64], [64, 3]]
)


flow = workflow.RLWorkFlow(configs)
flow.run_episodes()

Here you will be able to see the game, and a very nice dashboard with all the runs and configurations and nice graphs. Stay tuned with rewards.ai for further updates, documentation and examples.