Alongside the release of the StarCraft II Learning Environment (SC2LE) Blizzard Entertainment and DeepMind also released a set of minigames. These minigames provide the opportunity to investigate certain tasks of the full Game in isolation.
Starcraft 2 mini-Games have various purposes which can be categorized based on how they're crafted. They are broadly grouped into three categories: single-player scenarios, multi-player scenarios and live melee.
Single-player scenarios involve a single agent performing some tasks against a built-in bot or .SC2Map with pre-scripted elements. These elements narrow the scope of the game to a small subset of actions that would normally be available in a full game (live melee). Often times these elements hard code a scoring algorithm into the map that induces agent behavior using predefined rewards. Training an agent in these kinds of scenarios leads to a general understanding of how to perform the intended skill. It's main advantage and primary purpose is to induce an agent to rapidly acquire basic game-playing skills.
These scenarios involve multiple agents performing a small subset that would normally be available in the game. Unlike single-player scenarios, the objective of the scenario is not encoded in the map file. Agent performance is determined separately from the map.
In these scenarios, each agent controls their own units against the other agent's units. Not only does this induce specific behaviors intended by the scenario, it allows the agent to experience different play styles of various opponents. In this way, over-training to a specific play style can be avoided by varying the opponent agents and allow the agent to be more adaptive to variations in the scenario, or similar situations in a real game, which can often occur if only training single-player scenarios.
In other words, the purpose of multi-player scenarios is to induce the agent to not only learn the scenario in a specific, predefined way but also learn how the opponent can subtly vary the scenario with varying outcomes.
The primary advantage over Live Melee is that these scenarios can be repeated in short order and enable an agent to learn specific scenarios very quickly. The learning rate for the specific skill (in time spent) is faster than playing many games that may only encounter that situation relatively few times and therefore not adequately train the agent to handle the particular, nuanced situation it encountered.
The primary advantage of live melee versus single-player or multi-player scenarios is to model the make available full range of actions of an entire game. By playing a full game, an agent is expected to apply its acquired skills collectively as it selectively deems appropriate to be able to play the full-scale match. Thus the purpose of live melee is to test the aggregate of all its accumulated skills to determine its overall proficiency at playing Starcraft 2.
While the requirements for any given mini-game can vary, each mini-game must minimally be defined by a .SC2Map. Once the .SC2Map is selected and loaded, the objective of the mini-game may become apparent as the agent interacts with what's made available in the map and perhaps the rules that the map enforces.
- Joint paper between Blizzard Entertainment and DeepMind that besides others describes the initial set of mini-games and reports initial baseline results.
- Talk about mini-game design
- Starcraft2 Maps with Simple, Universal Retrieval
- StarcraftGym - Provides leaderboards of minigame learning agents
- PySC2 OpenAI Gym Environments