Dataset Cards - OG MARL

3m - Download

Metadata

Environment name	Version	Agents	Action type	Observation size	Reward type
SMAC (v1)	SMAC V1, from OxWhiRL	3	Discrete	[30]	Dense

Generation procedure for each dataset

A QMIX system was trained to target level of performance. The learnt policy was then rolled out to collect approximately 250k transitions. An epsilon greedy policy with eps=0.05 was used. This procedure was repeated 4 times and the data was combined.

Summary statistics

Uid	Episode return mean	Min return	Max return	Transitions	Trajectories	Joint SACo
Poor	4.69 ± 2.14	0.00	20.00	997370	48779	0.81
Medium	9.96 ± 6.06	0.00	20.00	995313	41619	0.85
Good	16.49 ± 5.92	0.00	20.00	996366	43559	0.80

8m - Download

Metadata

Environment name	Version	Agents	Action type	Observation size	Reward type
SMAC (v1)	SMAC V1, from OxWhiRL	8	Discrete	[80]	Dense

Generation procedure for each dataset

A QMIX system was trained to target level of performance. The learnt policy was then rolled out to collect approximately 250k transitions. An epsilon greedy policy with eps=0.05 was used. This procedure was repeated 4 times and the data was combined.

Summary statistics

Uid	Episode return mean	Min return	Max return	Transitions	Trajectories	Joint SACo
Poor	5.28 ± 0.56	0.00	7.62	995144	20629	0.64
Medium	10.14 ± 3.34	0.00	20.00	996501	39208	0.96
Good	16.86 ± 4.33	0.19	20.00	997785	30638	0.86

5m_vs_6m - Download

Metadata

Environment name	Version	Agents	Action type	Observation size	Reward type
SMAC (v1)	SMAC V1, from OxWhiRL	5	Discrete	[55]	Dense

Generation procedure for each dataset

A QMIX system was trained to target level of performance. The learnt policy was then rolled out to collect approximately 250k transitions. An epsilon greedy policy with eps=0.05 was used. This procedure was repeated 4 times and the data was combined.

Summary statistics

Uid	Episode return mean	Min return	Max return	Transitions	Trajectories	Joint SACo
Poor	7.45 ± 1.48	0.00	20.00	934505	45501	0.85
Medium	12.62 ± 5.06	0.00	20.00	996856	39284	0.87
Good	16.58 ± 4.69	0.00	20.00	996727	36311	0.84

2s3z - Download

Metadata

Environment name	Version	Agents	Action type	Observation size	Reward type
SMAC (v1)	SMAC V1, from OxWhiRL	5	Discrete	[80]	Dense

Generation procedure for each dataset

A QMIX system was trained to target level of performance. The learnt policy was then rolled out to collect approximately 250k transitions. An epsilon greedy policy with eps=0.05 was used. This procedure was repeated 4 times and the data was combined.

Summary statistics

Uid	Episode return mean	Min return	Max return	Transitions	Trajectories	Joint SACo
Poor	6.88 ± 2.06	0.00	13.61	996418	9942	0.96
Medium	12.57 ± 3.14	0.00	21.30	996256	18605	0.98
Good	18.32 ± 2.95	0.00	21.62	995829	18616	0.98

3s5z_vs_3s6z - Download

Metadata

Environment name	Version	Agents	Action type	Observation size	Reward type
SMAC (v1)	SMAC V1, from OxWhiRL	8	Discrete	[136]	Dense

Generation procedure for each dataset

A QMIX system was trained to target level of performance. The learnt policy was then rolled out to collect approximately 250k transitions. An epsilon greedy policy with eps=0.05 was used. This procedure was repeated 4 times and the data was combined.

Summary statistics

Uid	Episode return mean	Min return	Max return	Transitions	Trajectories	Joint SACo
Poor	5.90 ± 2.22	0.19	11.93	996474	17807	0.96
Medium	10.69 ± 1.49	0.00	17.67	996699	18866	0.97
Good	16.56 ± 3.72	6.30	24.46	996528	7315	0.97

terran_5_vs_5 - Download

Metadata

Environment name	Version	Agents	Action type	Observation size	Reward type
SMAC (v2)	SMAC V2, from OxWhiRL	5	Discrete	[82]	Dense

Generation procedure for each dataset

A QMIX system was trained to target level of performance. The learnt policy was then rolled out to collect approximately 250k transitions. An epsilon greedy policy with eps=0.05 was used. This procedure was repeated 4 times and the data was combined.

Summary statistics

Uid	Episode return mean	Min return	Max return	Transitions	Trajectories	Joint SACo
Replay	10.05 ± 5.84	0.00	36.34	898164	17958	1.00
Random	2.43 ± 1.73	0.00	16.18	1500000	37874	0.91

terran_10_vs_10 - Download

Metadata

Environment name	Version	Agents	Action type	Observation size	Reward type
SMAC (v2)	SMAC V2, from OxWhiRL	10	Discrete	[162]	Dense

Generation procedure for each dataset

A QMIX system was trained to target level of performance. The learnt policy was then rolled out to collect approximately 1m transitions. An epsilon greedy policy with eps=0.05 was used.

Summary statistics

Uid	Episode return mean	Min return	Max return	Transitions	Trajectories	Joint SACo
Replay	6.32 ± 3.62	0.00	23.01	749850	13588	1.00

zerg_5_vs_5 - Download

Metadata

Environment name	Version	Agents	Action type	Observation size	Reward type
SMAC (v2)	SMAC V2, from OxWhiRL	5	Discrete	[82]	Dense

Generation procedure for each dataset

A QMIX system was trained to target level of performance. The learnt policy was then rolled out to collect approximately 1m transitions. An epsilon greedy policy with eps=0.05 was used.

Summary statistics

Uid	Episode return mean	Min return	Max return	Transitions	Trajectories	Joint SACo
Replay	7.34 ± 3.60	0.00	24.00	863281	23294	1.00

2halfcheetah - Download

Metadata

Environment name	Version	Agents	Action type	Observation size	Reward type
MAMuJoCo	V1.1, Mujoco v210	2	Continuous	[13]	Dense

Generation procedure for each dataset

A MATD3 system was trained to target level of performance. The learnt policy was then rolled out to collect approximately 250k transitions. Gaussian noise with standard deviation of 0.2 was added to the action selection. This procedure was repeated 4 times and the data was combined.

Summary statistics

Uid	Episode return mean	Min return	Max return	Transitions	Trajectories	Joint SACo
Poor	400.45 ± 333.96	-191.49	905.03	1000000	1000	1.00
Medium	1485.00 ± 469.14	689.43	2332.17	1000000	1000	1.00
Good	6924.11 ± 1270.39	803.12	9132.25	1000000	1000	1.00