Skip to content

Using deep expected sarsa with tensorflow to solve the lunar lander problem with hyperparameter tuning and results analysis

Notifications You must be signed in to change notification settings

mokeddembillel/Lunar-Lander-Deep-Expected-Sarsa

Repository files navigation

Lunar-Lander-Deep-Expected-Sarsa

This project is about using deep expected sarsa with tensorflow to solve the lunar lander problem with hyperparameter tuning and result's analysis

These points are important read them please:

  • all data and models are in results folder

  • folder tests is just for you if you want to test something

  • if you want to plot the data, you don't need tensorflow just python and matplotlib

  • you need TensorFlow 2.3 for it to work

the code is structured this way:

  • imports

  • replay buffer class

  • expected sarsa network

  • softmax and argmax helper functions

  • agent class

  • lunarlander class

  • loading, parsing plotting helper functions

  • run experiment function for testing egreedy and softmax

  • run experiment function for testing batch size and replay steps

  • run experiment normal run

  • setting all learning parameters, data and the call for the run experiment function

  • cell for loading data and defining the loaded variable (run it before trying to plot or test the agent)

  • cell to compute averages

  • cell to plot same : plot all data (rewards, loss, episode steps) for each type of tests

  • cell for choosing best batch size and number of replay steps

  • cell for choosing best softmax tau and step-size

  • cell for choosing best e-greedy epsilon and step-size

  • INFO CELL (IMPORTANT READ IT PLEASE)

  • cell to load and reconstruct model

  • cell to plot data of best model

  • cell to test the agent

  • thank you :)

About

Using deep expected sarsa with tensorflow to solve the lunar lander problem with hyperparameter tuning and results analysis

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published