Grunnleggende konsepter
Incorporating an external Theory of Mind (ToM) model to predict the future actions of AI agents can enhance human understanding of agent intentions and improve the efficiency of human-AI collaboration.
Sammendrag
The paper proposes a two-stage paradigm to assist humans in human-AI coordination tasks. In the first stage, a ToM model is trained on offline trajectories of the target AI agent to learn to predict its future actions. In the second stage, the trained ToM model is utilized during the real-time human-AI collaboration process to display the predicted future actions of the AI agent, helping the human better understand the agent's intentions.
The key highlights of the paper are:
- The proposed paradigm does not require any prior knowledge of the environment or the AI agent's algorithm, making it compatible with general deep reinforcement learning (DRL) scenarios.
- The ToM model is trained as a standalone component and can be regarded as a third-party assistant, without affecting the behavior of the target AI agent.
- The authors implement a transformer-based ToM model and develop an extended Overcooked environment to support the visual presentation of agent intentions.
- Extensive experiments are conducted with two types of DRL agents (self-play and fictitious co-play) across multiple layouts, demonstrating that the ToM model can significantly improve the performance and situational awareness of human-AI teams.
- The user assessment reveals that the ToM model can enhance the human's satisfaction with the predictor and their understanding of the AI agent's intentions, leading to better collaboration efficiency.
Statistikk
The paper reports the following key metrics:
Average rewards of human-AI teams with and without the ToM model across different layouts and agent types.
Prediction accuracy of the ToM model on the offline test set and in the real human-AI collaboration experiments.
Sitater
"Our paradigm does not require any prior knowledge of the environment, the prediction is at the action level, ensuring its availability in general DRL scenarios."
"The ToM model is trained from offline data and can be regarded as a complete post-process that has no effect on the behavior of the target agent, providing compatibility for all DRL algorithms."
"In our paradigm, the agent can be regarded as a gray box or even a black box, which establishes the potential of the ToM model to be a third-party assistant for practical applications."