clj-gym

Clojure bindings for OpenAI's gym-http-api server.

Installation

With Leiningen/Boot:

[mthomure/clj-gym "0.1.1-SNAPSHOT"]

Usage

The clj-gym.core namespace follows OpenAI's API description.

> (require '[clj-gym.core :as g])

> (g/create "CartPole-v0")
"2f7eef5a"

> (g/envs)
{"2f7eef5a" "CartPole-v0"}

> (g/reset "2f7eef5a)
{:observation [0.023710455775827657 0.019351021639290936 0.011374305124354243 -0.016630857354495036]}

> (g/step "2f7eef5a" 0)
{:done false, :info {}, :observation [0.024097476208613477 -0.1759321880637575 0.011041687977264342 0.2796189824583605], :reward 1.0}

Get an idea of valid actions and observations using:

> (g/action-space "2f7eef5a")
{:n 2, :name "Discrete"}

> (g/observation-space "2f7eef5a")
{:high [4.8 3.4028234663852886E38 0.41887902047863906 3.4028234663852886E38], :low [-4.8 -3.4028234663852886E38 -0.41887902047863906 -3.4028234663852886E38], :name "Box", :shape [4]}

Talking to a remote server instance? Use the *url* dynamic variable.

> (binding [g/*url* "http://HOST:PORT"] (g/envs))

Running the Server

The server is available from DockerHub. To render a video and access it in the monitor files:

docker run -p 5000:5000 -v /tmp:/tmp mthomure/openai-gym-server

Then run in a repl:

(let [id (g/create "CartPole-v0")]
  (g/start-monitor id "/tmp/monitor" :video-callable true)
  (g/reset id)
  (g/step id 0 :render? true)
  (g/close-monitor id))

The monitor files should be in /tmp/monitor on the host filesystem. Note the use of :video-callable and :render?, above.

Warning: If you start-monitor with an existing directory, you'll get an HTTP 500 error. In this case, use :force? or :resume?.

Rebuilding the Docker Image

cd docker/
docker build -t mthomure/openai-gym-server .
docker login
docker push mthomure/openai-gym-server