tyluhhhh/rl_course_vizdoom_doom_deathmatch_bots

This model is trained to run around a map trying to kill bots in the videogame Doom. It has the functions of walking around on a 2d plane, with 360 degrees of mobility, as well as having a 360 degree camera movement function.

A(n) Factory sample model trained on the doom_deathmatch_bots environment.

This model was trained using Sample-Factory 2.0: https://github.com/alex-petrenko/sample-factory.

My process:

goal:

After initially making a model that succesfully was able to strafe left and right to shoot bots, i decided to make an even more complicated model that would instead move around the map with 360 camera movement and 360 movement to kill bots that way.

challenges:

I may have overestimated the AI's capabilities and its success rate with harder challenges. Adding on the second dimension of lateral movement as well as tacking on 3 dimensional camera movement proved to be simply too much for training of this caliber (around 50 minutes on my laptop using the T4 GPU provided by Colab) and the model was unable to perform well, as seen in the video.

improvements:

I can try to run a longer and more complex training with a more powerful computer (my home pc) since this will increase the accuracy of the ai. I could have also instead of adding camera movement and 2D movement, only add one or the other. For example, i could have pivoted to having the AI swivel its camera on a flat plane to shoot enemies, then add the aspects of movement in. In summary, my overconfidence in the initial model's success and my overestimation of such simple training for an AI led to a 20 minute video of an AI doing circles and randomly shooting.

It is worth noting however, since the AI only bases its next move off of one frame of the game its fed, the AI bot can be seen going in circles, which im assuming is it trying to find an enemy to shoot.

Downloading the model

After installing Sample-Factory, download the model with:

python -m sample_factory.huggingface.load_from_hub -r tyluhhhh/rl_course_vizdoom_doom_deathmatch_bots

Using the model

To run the model after download, use the enjoy script corresponding to this environment:

python -m <path.to.enjoy.module> --algo=APPO --env=doom_deathmatch_bots --train_dir=./train_dir --experiment=rl_course_vizdoom_doom_deathmatch_bots