LLM_Negotiations

This repository leverages Ollama.

1st install Ollama through https://ollama.com

2nd Run ollama in your terminal through

ollama run llama3

3rd Paste this line of code in yout terminal to create a tailored LLM named "reader"

ollama create reader -f ./V2_Reviewing_potential_offers/Modelfile

Finally, Run the version of the you Pyhton script prefer (Negotiaton_LLM_VX.py)

V4 also requires the input: ollama create constrain_reader -f./V4_Understanding_Constrain/Modelfile_reader_of_constrains

Arquitecture Iterations

V0:

• System Prompt: Negotiation Context and Principles
• Introduction Prompt: Max length, Be Proactive
• Follow up Prompt: Conversation History, Attention to last Message

V1:_ (Changes)_

• System Prompt: Rules, Combined Payoff (Price & Quality)
• Follow-up User Prompt: Instruction based (1st 2nd and 3rd)

V1.5:

• System Prompt: Separate Payoffs one for Price and one for Quality

V1.5.1:

• System Prompt: Is loaded by using Modelfile principles with temperature of 0.2
  • https://github.com/ollama/ollama/blob/main/docs/modelfile.md

V2:

• NEW AGENT: Main task is understand the offer and output the Price and Quality
• Rule-Based on User: If-Else Statement evaluates the profitability and prepares tailored prompts
• The V1.5 Agent: Receives the (Human Designed) profit evaluation to reply to the counterpart accordingly.

V3:

• Rule-Based on both User and Bot: If-Else Statement evaluates the profitability of the offer of the user and bot and prepares tailored prompts

V4:

• Totally NEW Bot Led: Understand the constrain of the counterpart first
• Actively ask for the user Ideal P & Q combination: Improved check of best offer within the bot loop (now it contrasts against the best user offer)
• Directly calculate the user profitability and own: Stored in separate dictionaries (Next step is pareto efficient bidding strategy)

V4.5:

• Pareto Efficieng Bidding Strategy: Leverage the profit function to send offers that are maximizing the bot profits while finding a common balanced ground.
• Bot Prompts Adapted to BOTH Roles: Now the prompting is dynamic based on the bot role and their constrain value.
• Offers Acceptance Is Greedy or Regular:
  • Greedy: Bot will only accept offers that yield a profit higher or equal to the maximum profit in the pareto efficient frontier.
  • Regular: Bot will only accept offers that yield a profit higher or equal to the minimum profit in the pareto efficient frontier.
• This version requires the ollama Models constrain_reader and reader, check OllamaCode.txt to install them.