StarCraft Multi-Agent Challenge (SMAC)#

What is SMAC#

The StarCraft Multi-Agent Challenge (SMAC) is a benchmark problem for deep multi-agent reinforcement learning. Focusing on partially observable, cooperative multi-agent learning, SMAC is based on the real-time strategy game StarCraft II and provides a diverse set of micromanagement scenarios for agents to learn. The challenge emphasizes the lack of standardized benchmarks in cooperative multi-agent RL and introduces SMAC as a solution.

How is SMAC Designed#

SMAC scenarios are carefully designed to evaluate how individual agents learn to coordinate to solve complex tasks. Each scenario is a confrontation between two armies, with varying initial positions, quantities, and unit types. Sometimes, there are also terrain features like high ground or impassable areas. The goal is to maximize the win rate, the ratio of games won to games played.

Why We Choose SMAC#

The scenarios cover a wide range of combat combinations and tactical challenges, providing a rich testing platform for multi-agent reinforcement learning algorithms. Through these detailed parameters, researchers can precisely control and understand the characteristics of each scenario, better evaluating and optimizing their algorithms.

Information of SMAC Scenarios#

Below is a detailed description of the scenarios, including the number of agents, enemies, time limits, races, and other specific parameters.

SMAC Maps Information

Map Name

Difficulty

Agents

Enemies

Limit

Agent Race

Enemy Race

2m_vs_1z

Easy

2

1

150

T

P

2s3z

Easy

5

5

120

P

P

2s_vs_1sc

Easy

2

1

300

P

Z

3m

Easy

3

3

60

T

T

3s_vs_3z

Easy

3

3

150

P

P

3s_vs_4z

Easy

3

4

200

P

P

8m

Easy

8

8

120

T

T

so_many_baneling

Easy

7

32

100

P

Z

MMM

Hard

10

10

150

T

T

1c3s5z

Easy

9

9

180

P

P

bane_vs_bane

Easy

24

24

200

Z

Z

2c_vs_64zg

Hard

2

64

400

P

Z

3s5z

Hard

8

8

150

P

P

5m_vs_6m

Hard

5

6

70

T

T

3s_vs_5z

Hard

3

5

250

P

P

8m_vs_9m

Hard

8

9

120

T

T

25m

Hard

25

25

150

T

T

10m_vs_11m

Hard

10

11

150

T

T

27m_vs_30m

Super Hard

27

30

180

T

T

3s5z_vs_3s6z

Super Hard

8

9

170

P

P

6h_vs_8z

Super Hard

6

8

150

Z

P

MMM2

Super Hard

10

12

180

T

T

corridor

Super Hard

6

24

400

P

Z

Hint

In the naming of SMAC maps, certain letters represent specific unit types. Here’s an explanation of these abbreviations:

  • m: Marine (Terran infantry unit)

  • h: Hydralisk (Zerg ranged unit)

  • z: Zealot (Protoss melee unit)

  • c: Colossus (Protoss massive unit)

  • zg: Zergling (Zerg basic melee unit)

Additionally, the map information includes the three main StarCraft races:

  • T: Terran (Human race)

  • Z: Zerg (Insect-like race)

  • P: Protoss (Advanced alien race)

These units and races appear in various combinations across the maps, defining the characteristics and difficulty of each scenario.

Explanation of Parameters: - Agents: Number of agents in the scenario. - Enemies: Number of enemy units. - Limit: Maximum length of action sequence allowed. - Agent Race: Race of the agents (Terran, Protoss, Zerg). - Enemy Race: Race of the enemies (Terran, Protoss, Zerg).

For the full details of the scenarios and the research, please refer to the original paper.