MrMagnifico / cse3210-negotiation-agent-group-66

Negotiating Agent GrugAgent

GrugAgent is a modern negotiating agent based on the winner of the 2017 edition of the Automated Negotiating Agent Competition held by the Delft University of Technology. It has been made as the first assignment of the CSE3210 Collaborative Artificial Intelligence elective. It has been written by Valentijn van de Beek, William Narchi, and Emirhan Demir. This guide assumes that you are using the version hosted on github.com, otherwise the normal installation instructions for Python based agents apply. For the Tomcat server the same principles mostly apply, but the settings need to be set manually. These are described below but are not tested due to a bug with the SSL certificate used in the tomcat server.

Resources

Negotiating Agents (Catholijn M. Jonker, Koen V. Hindriks, Pascal Wiggers, Joost Broekens) Bayesian learning in negotiation (Dajun Zeng, Katia Sycara) Multi-Issue Automated Negotiations Using Agents(K. Chari, M. Agrawal)

Directory overview

• directories:
  • agents: Contains directories with the agents. The ponpokoagent/ponpoko directory contains the template for this assignment.
  • ponpokoagent/ponpoko/ponpoko.py is the file containing the code for the agent including the opponent modeling.
  • ponpokoagent/ponpoko/patterns.py contains the generation which provides the patterns. It contains all the code for the standard pattern, custom patterns and generating new patterns.
  • domains: Contains the domains which are problems over which the agents are supposed to negotiate.
  • utils: Arbitrary utilities (don't use).
• files:
  • grugagent.py: Main interface to test agents.
  • run_parallel.py: Command-line program to run tournaments in parallel and allows you to set the options for the agent using switches
  • requirements.txt: Python dependencies for your agent.
  • requirements_allowed.txt: Additional dependencies that you are allowed to use (ask TA's if you need unlisted packages).

Quickstart

• Create a python environment: python -m venv .py
• Load the python enviornment: source .py/bin/activate
• Install the packages: pip install --trusted-host tracinsy.ewi.tudelft.nl -r requirements.txt
• Run the default tournaments: python run_parallel.py

When the program finishes running you can find the output in results/[domain]. In that directory there is the general summary, a figure with the normal distribution (provided --matplotlib is given) and a JSON file detailing each run of the program. It is recommended that the files are parsed using a tool such as [jq](https://stedolan.github.io/jq/).

Commandline arguments

GrugAgent contains a large variety of options which control the usage and result of the agent. These can be changed manually in the code or, more conviently, by using the switches provided by run_parallel.py. You can always check what it accepts by passing it the -h switch, but a more full explanation is given below.

Options

• samples: Amount of times that each tournament is repeated. These are aggregated together at the end and are used to calculate interesting statistical data about the run.
• frequency: Number of turns before the agent switches to a different pattern. Corresponds to the patternChangeDelay setting parameter.
• fallback-tol: Defines the maximum difference between the chosen bid compared the median of all the bids. Corresponds to the fallbackBidUtilRange setting parameter.
• types: Allows the user to change between the different operating modes of PonPoko. They are described in detail below. This value is the same as the generatorType setting parameter and is a number between 1 and 5.
• domains: A list of integers between 0 and 9 which indicate which domains the tournaments are run on.
• agents: List of agents that are used as opponents in the tournament.
• matplotlib: Switch which generates a figure with the normal distribution defined by the output the aggregated samples.
• opponent-epsilon-higher: Change required before the opponent is classified as a hardliner. Called opponentEpsilonLower in the settings.
• opponent-epsiolon-lower: Change requried before the opponent is classified as a conceder. Corresponds to the opponentEpsilonLower setting parameter.

Agents available

These are the agents which are available to be used in the tournaments, more could be added easily by adding them to the agents/ directory.

• BoulwareAgent: Adjustable hardliner agent
• ConcederAgent: Conceder agent
• Hardlineragent: Hardliner agent
• LinearAgent: An agent that uses a linear formula to accept or reject bids
• RandomAgent: Randomly accepts or rejects bids
• StupidAgent: Always accepts the first bid
• TemplateAgent: Base agent provided by the university
• CustomAgent: Template agent provided by the university
• RandomParty: Version of the random agent
• PonPokoParty: Base PonPokoAgent

PonPoko operator modes

In this project the goal was to find improvements which would cause measurable better utility gain over the standard PonPokoAgent. To do so four additional modes are added which represent different modifications of the agent.

1. standard: Based directly on the ANAC 2017 paper and uses the base patterns
2. opponent: Version of Pon Poko with frequency opponent modeling
3. random: Randomly generates the parameter of the patterns
4. mutation: Uses mutation rules to change the operators used in the patterns
5. generation: Mutates the operators and uses random numbers for the parameters.

More information

More documentation can be found here. Part of this documentation was written for the Java version of GeniusWeb, but classes and functionality are left identical as much as possible.