Hands-On Reinforcement Learning for Games

Revisiting the LunarLander and beyond

Now with our solid example of DQN, we can move on to solve more difficult environments, like LunarLander. In this exercise, we set up the DQN agent to solve the LunarLander environment in order to compare our previous attempts with discretized SARSA:

1. Open the Chapter_6_DQN_lunar.py example, and note the change in the env_id environment ID and creation of the environment shown as follows:

env_id = 'LunarLander-v2'env = gym.make(env_id)

2. We also adjust a couple of the hyperparameters to account for the increased complexity of the environment:

epsilon_decay = 1000buffer_size = 3000neurons = 192

3. We increase epsilon_decay in order to encourage the agent to explore longer. Exploration is a trade-off we ...