Gym Skill

The AEA gym skill demonstrates how a custom Reinforcement Learning agent, that uses OpenAI's gym library, may be embedded into an AEA skill and connection.

Discussion

The gym skills demonstrate how to wrap a Reinforcement Learning agent in a skill. The example decouples the RL agent from the gym.Env allowing them to run in separate execution environments, potentially owned by different entities.

Preparation Instructions

Dependencies

Follow the Preliminaries and Installation sections from the AEA quick start.

Download the necessary directories into your working directory:

mkdir gym_skill_agent
svn export https://github.com/fetchai/agents-aea.git/trunk/examples

Install the gym and numpy library.

pip install numpy gym

Demo Instructions

Create the AEA

First, fetch the gym AEA:

aea fetch fetchai/gym_aea:0.26.5 --alias my_gym_aea
cd my_gym_aea
aea install

??? note "Alternatively, create from scratch:"

### Create the AEA

In the root directory, create the gym AEA and enter the project.

``` bash
aea create my_gym_aea
cd my_gym_aea
```

### Add the gym skill

``` bash
aea add skill fetchai/gym:0.21.6
```

### Set gym connection as default

``` bash
aea config set agent.default_connection fetchai/gym:0.20.6
```

### Install the skill dependencies

To install the `gym` package, a dependency of the gym skill, from PyPI run

``` bash
aea install
```

Set up the Training Environment

Copy the Gym Environment to the AEA Directory

mkdir gyms
cp -a ../examples/gym_ex/gyms/. gyms/

Update the Connection Configuration

aea config set vendor.fetchai.connections.gym.config.env 'gyms.env.BanditNArmedRandom'

Create and Add a Private Key

aea generate-key fetchai
aea add-key fetchai

Run the AEA with the Gym Connection

aea run

You will see the gym training logs.

Delete the AEA

When you're done, you can go up a level and delete the AEA.

cd ..
aea delete my_gym_aea

Communication

This diagram shows the communication between the AEA and the gym environment

    sequenceDiagram
        participant AEA
        participant Environment

        activate AEA
        activate Environment
        AEA->>Environment: reset
        loop learn
            AEA->>Environment: act
            Environment->>AEA: percept
        end
        AEA->>Environment: close

        deactivate AEA
        deactivate Environment

Skill Architecture

The skill consists of two core components: GymHandler and GymTask.

In the setup method of the GymHandler the GymTask is initialized, as well as its setup and execute methods called. The handler, which is registered against the GymMessage.protocol_id then filters for messages of that protocol with the performative GymMessage.Performative.PERCEPT. These messages are passed to the proxy_env_queue of the task.

The GymTask is responsible for training the RL agent. In particular, MyRLAgent is initialized and trained against ProxyEnv. The ProxyEnv instantiates a gym.Env class and therefore implements its API. This means the proxy environment is compatible with any gym compatible RL agent. However, unlike other environments it only acts as a proxy and does not implement an environment of its own. It allows for the decoupling of the process environment of the gym.env from the process environment of the RL agent. The actual gym.env against which the agent is trained is wrapped by the gym connection. The proxy environment and gym connection communicate via a protocol, the gym protocol. Note, it would trivially be possible to implement the gym environment in another AEA; this way one AEA could provide gym environments as a service. Naturally, the overhead created by the introduction of the extra layers causes a higher latency when training the RL agent.

In this particular skill, which chiefly serves for demonstration purposes, we implement a very basic RL agent. The agent trains a model of price of n goods: it aims to discover the most likely price of each good. To this end, the agent randomly selects one of the n goods on each training step and then chooses as an action the price which it deems is most likely accepted. Each good is represented by an id and the possible price range [1,100] divided into 100 integer bins. For each price bin, a PriceBandit is created which models the likelihood of this price. In particular, a price bandit maintains a beta distribution. The beta distribution is initialized to the uniform distribution. Each time the price associated with a given PriceBandit is accepted or rejected the distribution maintained by the PriceBandit is updated. For each good, the agent can therefore over time learn which price is most likely.

The illustration shows how the RL agent only interacts with the proxy environment by sending it action (A) and receiving observation (O), reward (R), done (D) and info (I).