Embedded Files
Here are some of my results from a Stanford class (234, Emma Brunskill) I took in 2023-4. All these use test RI environments provided by OpenAI's Gymnasium.
Here are some of my results from a Stanford class (234, Emma Brunskill) I took in 2023-4. All these use test RI environments provided by OpenAI's Gymnasium.
RL-pong-dqn.mp4Pong (DQN)
Pong (DQN)
Learning Pong implemented with DeepMind's deep Q-learning. My model's paddle is on the right. The cool thing about this is that it's learning directly from pixels! The input are the last four video frames of the gameplay, after preprocessing them: max-pooling between adjacent frames, converting to grayscale, cropping, rescaling, and stacking, so 210x160x3x4 becomes a 80x80x4 input.
The following use the REINFORCE policy gradient algorithm, with an optional neural network baseline to reduce variance:
The following use the REINFORCE policy gradient algorithm, with an optional neural network baseline to reduce variance:
RL-HalfCheetah-no-baseline.mp4Half-Cheetah
Half-Cheetah
Learns locomotion in most unlikely way.
RL-CartPole-baseline.mp4Cart Pole
Cart Pole
Learns a balancing act, by applying forces horizontally to the cart.
RL-InvertedPendulum-baseline.mp4Inverted Pendulum
Inverted Pendulum
Similar to the last one, but powered by the Mujoco physics simulator.
Page updated
Google Sites
Report abuse