critic_multiType: critic multi type, one for each agent, within type mean field, between type attention
critics_singleType_vanilla: critic single type, one for each agent, independent
critics_vanilla: try 1 for one virtual critic for each type, single type
critics multiType_virtual: from critic_multiType, 1 virtual for each type
critic_multiType2: from critic_multiType, use more systematic way for mean field, should deprecate critic_multiType if critic_multiType2 result (mean field over actions) good
critic_multiType2_noAtt: from critic_multiType2, within type mean field, between type no attention, input to one network



