8000 GitHub - megforr/open_ai: Range of RL projects using OpenAI gym package
[go: up one dir, main page]
More Web Proxy on the site http://driver.im/
Skip to content

megforr/open_ai

Repository files navigation

open_ai

Range of RL projects using OpenAI gym package

Autonomous Taxi

Goal: Have a taxi agent learn how to navigate a grid world to pick up and drop off a passenger using tabular q-learning

  • This notebook was created to teach coworkers about Q-learning

Lunar Lander

Goal: Teach a DQN agent to learn how to land a ship on a landing pad.

Experiments:

  • Memory size (memory): experiment_results/avg_reward_dqn_11_23_2021_13_04.png


Observations: Nearing 250 episodes, agents trained with both 100k and 1E6 experience replay buffer size observed better 52CC rolling average rewards. Nearly double 1E7 memory size.

alt text

  • Learning rate (alpha): experiment_results/avg_reward_dqn_11_25_2021_01_53.png


Observations: Agent learned in a more stable fashion at alpha=0.001. Maneuvers created by agent learning at 0.01 seemed much more risky (swinging wildly from side to side) opposed to the conservative upright landing style of the 0.001 agent.

alt text

  • Discount rate (gamma): experiment_results/avg_reward_dqn_11_24_2021_20_27.png


Observations:

alt text

Note: Best results would be observed if multiple experiments were performed and variance was calculated.

About

Range of RL projects using OpenAI gym package

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published
0