What is vizdoom_clone_project_vibecoded?

This project trains a neural network to play the classic game DOOM using reinforcement learning on top of ViZDoom, a research platform that lets AI agents control DOOM through code. The AI does not use a language model to decide what to do in the moment. Instead it uses a convolutional neural network policy trained with a reinforcement learning algorithm called PPO, learning directly from the pixels of the game screen, since DOOM requires a decision roughly every 30 milliseconds, far faster than a text based AI model could respond. The project covers all 12 of ViZDoom's built in training scenarios, such as a simple room with one monster, a corridor with enemies, and health gathering tasks, as well as full original DOOM and DOOM II levels using any map. Some scenarios, like the basic room and a monster filled corridor, are already trained and perform well, while others are implemented but still need training runs. Free versions of the DOOM game files are bundled in, and the project can automatically use your own copy of the original game files if you own them and place them in the right folder. Because many of the built in scoring rules only weakly reward good play, the project adds optional bonus rewards on top, such as extra reward for exploring new areas, picking up weapons, dealing damage, or gaining health, with different combinations turned on for different scenarios. The project also includes tools for looking inside the trained network to understand what it has learned to pay attention to, using several established visualization techniques borrowed from academic computer vision research. These generate images showing which parts of the screen most influenced the AI's decisions, such as highlighting the monster in a fight scenario, and can be run against a trained model with a live game frame.