MoveToBeacon

Introduction

Implementation of an Artificial Intelligent Agent that learns to play the StarCraft II minimap MoveToBeacon by itself using only raw pixels.

Environment

In this project PySC2 is used. PySC2 is a Python-based interface for communicating with the game engine.

One important argument of the environment is step_mul. It defines the number of game steps per observation. With a screen size of 64 x 64 pixels and a step_mul of 3 the player moves at a speed of about 1 pixel per observation.

MoveToBeacon

MoveToBeacon is a map with 1 Marine and 1 Beacon. The Marine earns a reward of 1 for reaching the Beacon. If the Marine doesn't reach the Beacon, the Beacon disappears after 120 seconds (we call this time frame an episode). After one episode the environment is reset. For a perfect player it's possible to receive a total reward of about 30 after 120 seconds.

Run

Reinforce

$ python -m reinforce.run

A2C

$ python -m a2c.run

PPO

$ python -m ppo.run

Results

Reinforce

A2C

PPO