Minimal Info (Agent does not do well with massive prompt)

[2549486373]

We only want to give minimal information

• No history dialogue is given

• Somewhat like MCTS, keep one running statistic, one attitude judger to decide what is agent's attitude based on dialogue history, no historical information is passed.

• Keep the information given to execution proposition agents as minimal as possible.

• Give previous stat-action history (can give or not, even Markovian style, give past one, do AB testing)

• [x] #34

• [x] #33

• [x] #38

Previous dialogue give way too many information and overwhelms the agents. We try to keep minimal information flow.

[KevinBian107]

We can also do "banning" task depending on the attitude of the local agent, this should be judged by the HCA agent.

[KevinBian107]

Ensure format for the attitude info given to the attitude agent is the same

[KevinBian107]

Give only individual agent their attitude