CCF 2023 Taxing AI Competition

This repo provide the source code for the CCF 2023 Taxing AI Competition

Update!

• 2023.10.12: In the government track, we take the sigmoid of the reward at each time step. This is to avoid the superiority of policies terminating the episode early.

• 2023/10/09: Please note that the gov_task hyperparameter in the n4.yaml config file should be "gdp_gini". Our evaluation backend is constantly evaluating the gdp_gini of your submission according to the competition description.

Multi-Agent Game Evaluation Platform --- Jidi (及第)

Jidi supports online evaluation service for various games/simulators/environments/testbeds. Website: www.jidiai.cn.

A tutorial on Jidi: Tutorial

Environment

The competition adopts a Taxing simulator TaxingAI. A brief description can be found on JIDI. A complementary document is also presented in docs.

The game contains four roles, two of which are controllable. They are:

• A Government (controllable)
• N Households (controllable)
• Firm
• Bank

Government try to adjust tax parameters and the ratio of government spending to GDP, in order to optimize GDP and Gini index. The evaluation metric is the sum of sigmoid reward at each time step.

Each household try to balance saving ratio and working time, in order to optimize personal utility. The evaluation metric is the sum of individual at each time step.

Paper: TaxAI: A Dynamic Economic Simulator and Benchmark for Multi-Agent Reinforcement Learning

Quick Start

You can use any tool to manage your python environment. Here, we use conda as an example.

conda create -n taxingai-venv python==3.9  #3.10
conda activate taxingai-venv

Next, clone the repository and install the necessary dependencies:

git clone https://github.com/jidiai/Competition_TaxingAI.git
cd Competition_TaxingAI
pip install -r requirements.txt
conda install numba==0.58.0

Finally, run the game by executing:

python run_log.py

Navigation

|-- Competition_OvercookedAI               
	|-- agents                              // Agents that act in the environment
	|	|-- random                      // A random agent demo
	|	|	|-- submission.py       // A ready-to-submit random agent file
	|-- env		                        // scripts for the environment
	|	|-- config.py                   // environment configuration file
	|	|-- taxing_gov.py               // The environment wrapper for taxing_gov env	
	|   |-- taxing_household.py             // The environment wrapper for taxing_household env	      
	|-- utils               
	|-- run_log.py		                // run the game with provided agents (same way we evaluate your submission in the backend server)

How to test submission

• You can train your own agents using any framework you like as long as using the provided environment wrapper.

• For your ready-to-submit agent, make sure you check it using the run_log.py scrips, which is exactly how we evaluate your submission.

• run_log.py takes agents from path agents/ and run a game. For example:

python run_log.py --my_ai "random" --opponent "random"

set both agents as a random policy and run a game.

• You can put your agents in the agent/ folder and create a submission.py with a my_controller function in it. Then run the run_log.py to test:

python run_log.py --my_ai your_agent_name --opponent xxx

• If you pass the test, then you can submit it to the Jidi platform. You can make multiple submission and the previous submission will be overwritten.