PSO-Tetris-AI

This is a Term Project for CS3243 - Introduction to Artificial Intelligence at the National University of Singapore.

Getting Started

These instructions will get you a copy of the project up and training the AI on your local machine. More details will be added in the future.

Clone the project and compile.

git clone https://github.com/OscarWang114/PSO-Tetris-AI.git
cd PSO-Tetris-AI/src
javac *.java jswarm_pso/*.java

Running the AI

Simply run PlayerSkeleton to see the AI in action. A GUI tetris grid will be displayed for monitoring AI's playing process.

java PlayerSkeleton

Training the AI

Create a tetris_log folder in your home directory for the AI to log its training progress/result as a .csv file.

mkdir ~/tetris_log

Start training the Tetris AI. Note that you need to supply three arguments in the command for PSO. x1: Number of particles x2: Number of iterations x3: Number of threads

java PSO x1 x2 x3

For example: java PSO 30 100 3 uses 30 particles and 100 iterations to train the data. The 30 particles are distributed among 3 threads, so each thread gets 10 particles and processes them sequentially. However, the 3 threads are executed in parallel, which can theoretically reduce the computational time by a factor of 1/3 compared to java PSO 30 100 1.

Results

The best weights we obtained after PSO training is

Heuristic function	Weight value
Rows Cleared	-1.5027068
Total Height	-0.11106747
Maximum Height	0.43475921
Height Differences	0.80292475
Number of Holes	3.0
Deepest Well	1.7848895
Column Transition	3.0
Row Transition	0.87486663

The agent was able to clear more than 600K lines on average across 5 games without using the "one-piece lookahead" feature.

After enabling "one-piece lookahead", the agent cleared more than 1.4M lines on average across 10 games before we forced it to stop due to time limitations. We believe it could have cleared more than 2, 3M lines if we had let it continue playing for a longer time.

Automated Training on Sunfire using Cron

We used the Sunfire Server provided by the National University of Singapore and Cron to automate the training phase. A guide on using Sunfire and Cron written by OscarWang114 can be found here.

Built With

Java - The language used
Cron - The job scheduler used for training

Authors

Oscar Wang - Java Coding (heurstics, PSO training, multithreading), Cron Job executions, and Report Writing - OscarWang114
Vivian Nguyen - Java Coding (heurstics) and Report Writing - vivianmnguyen
Meriem Hrittane - PSO Studying, Researching, and Report Writing- mhrittane
Raphael Massart - Report Writing - RaphaelMassart
Aishwarya George - Report Writing - ash-g777

Acknowledgments

JSwarm-PSO library and its author Pablo Cingolani (pcingola@users.sourceforge.net).

OscarWang114 / PSO-Tetris-AI