notesum.ai

Published at November 21

Learning to Cooperate with Humans using Generative Agents

cs.AI

cs.MA

Released Date: November 21, 2024

Authors: Yancheng Liang¹, Daphne Chen¹, Abhishek Gupta¹, Simon S. Du¹, Natasha Jaques¹

Aff.: ¹University of Washington

Arxiv: http://arxiv.org/abs/2411.13934v1

Refer to caption

Agent	Training data source	Counter circuit	Multi-strategy Counter
FCP	FCP-generated population	26.24 $\pm$ 0.96	34.35 $\pm$ 1.47
FCP + GAMMA	FCP-generated population	57.87 $\pm$ 1.35	65.36 $\pm$ 1.24
CoMeDi	CoMeDi-generated population	56.87 $\pm$ 1.43	27.11 $\pm$ 1.46
CoMeDi + GAMMA	CoMeDi-generated population	72.17 $\pm$ 1.61	34.72 $\pm$ 1.65
MEP	MEP-generated population	76.19 $\pm$ 1.89	64.02 $\pm$ 2.07
MEP + GAMMA	MEP-generated population	81.10 $\pm$ 2.03	88.30 $\pm$ 2.51
PPO + BC	human data	53.51 $\pm$ 1.37	85.26 $\pm$ 2.28
PPO + BC + GAMMA	human data	59.67 $\pm$ 1.35	77.53 $\pm$ 2.00
GAMMA HA	MEP pop. $+$ human data	91.11 $\pm$ 2.96	93.09 $\pm$ 3.19