wernhervondata/ai-doom

Simple agent training wrapper for Vizdoom environment.

ai-doom's Issues

Agent for defending the center

Create an agent for defend the centre scenario, train it and create a movie how it performs - as a result share all the movies on the LinkedIn.

Do it on the base of the two input agent.

Improvements 1

• Create a ReplayMemory with random moves, then start training
• Start using CUDA

Use LSTM

Change the full connected layer from the linea neurons to the LSTM cells, so network might find correlations between the stages.

Extend the full connected layer with additional data from episodes

The additional numbers like health, ammo count etc. provide a really useful information about the environment - find a way how to extend the fully connected layer both with the information from the pooling and these numbers.

Load trained models and play them in episodes

Create a model loader to be able to show how the agent performs.

Improve player

Make player more configurable to avoid a lot of changing in the code between running the scenarios

Serialize ReplayMemory

Make the repo public

In order to make this repo public write a neat README file, where you mention both Vizdoom and SuperdataScience team

Extend the linear layer with additional features

Apart of health, add delta of health, step number and other features if you find them crucial

Record episodes

Check if it's possible to record episodes on Windows, and if yes - implement it

Fix rocket scenario training

The rocket basic scenario is not working right now due to different way of reading the images (don't know why, but the channels are different there).

Use ViZDoom with eligibility trace and cnn

As the course Doom environment is working anymore - prepare a better solution with the ViZDoom

Train with colab code

Check the colab code from the Udemy and compare how the network wrapped around gym looks like and search for potential bugs in my code