Bisca Agent Strategy Intelligent System, a multi-agent platform for implementing AI-driven strategies in the game of Bisca.

This repository hosts our project for the course of Autonomous Agents and Multi-Agent Systems (AASMA) @ Instituto Superior Técnico for the year of 2022/2023. A paper is also included, which provides a detailed overview of the project, its implementation, and the obtained results.

• José Almeida (IST1105793)
• Pedro Moço (IST105773)
• Tiago Teodoro (IST1105720)

This project is written in Python and makes use of a few external modules for purposes of UI and statistical plotting. Python 3.11 or newer is recommended.

In order to install all required dependencies, perform the following command at the project root.

pip install -r requirements.txt

Note: Switch pip for pip3 if the former isn't registered in your $PATH.

In order to interact with the platform via user-interface, you can launch it from the root of the project as follows.

./basis.py

Note: In case this doesn't work, prefix it with python or python3.

The game engine offers a comprehensible interface, a simple example of using the internal engine is provided at example.py.

simulate.py is an included script that allows you to run simulations over the BASIS Multi-Agent Platform. It provides a command-line interface to specify simulation parameters and visualize the results.

The simulation framework gathers and exposes metrics for each player such as total wins/draws/losses, win-rate, average points per game, and highest points obtained in a single trick.

Note: Simulation doesn't support Human players, as its grand objective is to obtain data relative to long-term comparison of agent types.

The script supports the following command-line arguments:

--iterations <int>      Number of simulations to run (default: 1000)
--delay <int>           Delay in seconds between each round (default: 0)
--player <type>         Player types to include (choices: all available agent types: RandomAgent, SimplyGreedyAgent, MinimizePointLossGreedyAgent, MPLGreedyTrumpSaveAgent, MPLGreedyTrumpBasedAgent, GreedyCountingAgent)
--graph                 Display graph of simulation results
--interpolate           Interpolate graph for smoother visualization (useful for large number of iterations)
--save                  Saves obtained metrics to a JSON file

Note: If, for some reason, the following instructions don't work for you, or you want to use a specific version of Python, prefix them with e.g. python or python3.

1. Run simulation with default parameters (1000 iterations with SimpleGreedyAgent, MinimizePointLossGreedyAgent, MPLGreedyTrumpSaveAgent, GreedyCountingAgent):

   ./simulate.py

2. Run simulation with specific player types:

   ./simulate.py --player RandomAgent SimpleGreedyAgent

3. Run 50 iterations and display the graph of simulation results:

   ./simulate.py --iterations 50 --graph --interpolate

   • The --graph option will display two sets of graphs in different tabs, one regarding overall player statistics, and another regarding per-iteration player statistics.

4. Run 10000 iterations with graph interpolation for smoother visualization, using specific agents, saving the obtained metrics to a JSON file:

   ./simulate.py --iterations 10000 --graph --interpolate --save --player RandomAgent SimpleGreedyAgent MinimizePointLossGreedyAgent MPLGreedyTrumpSaveAgent

Note The --interpolate option is only valid if --graph is specified.

If you're looking to analyze the obtained metrics, you can use the analysis.py script. It makes use of simulate.py constructs to compute various simulations for each group-size, for every possible combination of agents. It then aggregates each metric for each groups-zie and outputs the results to a JSON file.

To run it, simply execute the following command:

./analysis.py

Warning: This script may take a long time to run, as it performs 100000 simulations for each group-size, for each possible combination of agents (EXPERIMENT_SIMULATIONS is set to 100000, though you can change it to a lower value if you want to run it faster, although with less statistical confidence). If you're only looking for pre-computed metrics used to for the tables in the report, you can find them at report_analysis.json.