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