Deep Learning in ViZDoom - Defend the Center

Описание: Agent trained in ViZDoom environment (http://vizdoom.cs.put.edu.pl/) using visual input, current health points, current time, and ammo. Code that produced the behavior can be found here: https://github.com/ebonyclock/ai_vizdoom

The agent can only turn and shoot (limited ammo of 26 bullets), is rewarded for killing an attacker (1pt) and penalized for dying (-1pt). Apparent performance might not seem that great but really it is much better than I am. It is even sometimes able to score 26 and survive to the timeout (no penalty for death).