Train and Enjoy: REINFORCE CartPole
Last updated
Last updated
This tutorial will look at how to train an agent in SLM Lab, then use the automatically-saved model files to replay it in enjoy mode, using REINFORCE on CartPole.
REINFORCE is a very basic policy gradient algorithm, and we are training it on an easy environment. We will use the spec file at slm_lab/spec/benchmark/reinforce/reinforce_cartpole.json, in particular the reinforce_cartpole spec, and run it in train mode. The lab command is:
This will run a Trial with 4 Sessions of different random seeds to average the results. Wait for it to run until completion, which should take about 10-20 minutes. Meanwhile, check the metrics logged in the terminal. In particular, the total_reward and its moving average total_reward_ma (with a window of 100 episodes) should climb up gradually.
REINFORCE successfully trains on CartPole when the total_reward_ma reaches close to the maximum of 200, although a score of over 100 will do for this tutorial.
When the trial completes, all the metrics, graphs and data will be saved to a timestamped folder, let's say data/reinforce_cartpole_2019_11_18_233140/. Among other things, SLM Lab also automatically saves the final and the best model files in the model folder. The model files can be used for easy playback in enjoy mode.
We now look inside the model folder at data/reinforce_cartpole_2019_11_18_233140/model/. According to Lab Command the enjoy mode in SLM Lab also needs the prename, which is the spec name with trial and session. Let's use trial 0 session 2, so our prename is reinforce_cartpole_t0_s2.
The model files that corresponds to the final checkpoint is reinforce_cartpole_t0_s2. TO use this for enjoy mode, run the command using the saved spec file in the data folder, and the prename:
This will create a new Session by loading the saved spec file and the specified model. This will also render the environment like we saw in Quick Start, but only this time the REINFORCE agent which has the trained weights loaded will perform well immediately.
We can see this in the terminal from the logged metrics, with total_reward_ma starting at 200 (the maximum score) immediately:
Inside the model folder there are also variants of the model files saved with a suffix "ckpt-best". These files are saved from the best evaluation checkpoints (determined by total_reward_ma), and may or may not be equal to the final checkpoint. To run enjoy mode using them instead, simply supply the prename with "ckpt-best", like so:
Loading the best checkpoint model is especially useful when an algorithm's performance deteriorates over time.
In the next tutorial, we will dive into the spec file for an agent.
