Testing reliability of replay-based imitation for StarCraft

Abstract

For StarCraft, it's easy to download lots of replays from gaming portals. Using simple tools, it's possible to extract all the gaming events stored in the replays. At each frame, it can tell us the human player's decision making given game states. Instead of making hard-coded AIs, it's promising to imitate the human player's decision recorded in the replays. In this study, we propose to create an AI bot imitates human player's high-level decisions (attack or retreat) on a group of units from replays. As a first step, we tested the reliability of the imitation system using replays from portals. We reported the ratio of apparent mistakes from the imitation system and the way to reduce the error.