In this project, we trained a reinforcement learning agent to solve the HalfCheetah-v4 environment using mlpack's Twin Delayed Deep Deterministic Policy Gradient (TD3) implementation. The goal was to train the agent to perform efficiently in a complex control task.
We used the OpenAI Gymnasium Toolkit's GUI interface for training and testing our agent. This interface is provided through a distributed infrastructure (TCP API), which enables us to interact with the environment. You can find more details about this infrastructure here.
The full video output of the trained agent solving the environment can be found here
The code used to train the agent can be found here.
Here are the hyperparameters used for training the TD3 agent:
| Hyperparameter | Value |
|---|---|
| Training Steps | 150,000 |
| Step Size | 7e-4 |
| Target Network Sync Interval | 1 |
| Exploration Steps | 10,000 |
| Discount | 0.99 |
| Update Interval | 1 |
| Replay Buffer Size | 1,000,000 |