Shatranj-Chess Openings Analyzer with GenAI for Explainable AI

Table of Contents

• Project Overview
• Dataset
• Preprocessing
• Libraries-Used
• Data Pipeline Preparation
• Model Processing
• Demo

Project-Overview

The Chess Openings Analyzer project leverages machine learning (ML) techniques, including logistic regression, to analyze the success rates of various chess openings. By incorporating the latest advances in ML and data mining, such as deep learning algorithms and neural networks, this project offers several advantages over current state-of-the-art solutions.

Using machine learning algorithms like logistic regression and leveraging Gemini AI for a better understanding of the move suggested from the model we trained, the project predicts game results based on openings, ELO ratings, and other features in the dataset. Additionally, clustering techniques are employed to analyze the results of different openings, providing players with valuable insights into strategic decision-making.

Dataset

The dataset used in this project is sourced from Kaggle and offers a comprehensive collection of 6.25 million chess games played on lichess.org during July 2016. Each game is meticulously represented by a row, providing valuable insights into player strategies, game dynamics, and opening variations.

Preprocessing

A basic overview of the preprocessing steps applied to the Chess Games dataset is outlined below:

1. Find and Drop Missing Values: Identify and drop rows with missing values.
2. Drop NaNs: Remove any remaining NaN values.
3. Drop Unnecessary Columns: Remove the "White" and "Black" columns, as they only contain IDs that are not useful for analysis.
4. Encode Game Results: Encode game results into three categories: 1 for White win, 0 for Black win, and 2 for tie.
5. Drop UTC Date and Time: Remove UTC date and time columns as they are not required for analysis.
6. Categorize Events: Categorize and encode event types into fewer categories (e.g., Blitz, Blitz tournament).
7. Scale Elo Ratings: Scale White and Black Elo ratings for analysis.
8. Drop Rating Differences: Drop columns for White and Black rating differences as they provide limited value.
9. Create Opening DataFrame: Create a separate DataFrame for opening moves using the ECO code for mapping.
10. Drop Time Control: Temporarily drop the time control column for analysis, which may be useful for later.
11. Extract Individual Moves: Parse and extract individual moves from the AN column to analyze opening moves in detail.
12. Encode Termination Conditions: Encode termination conditions as "normal" (1) and "time forfeit" (0), while reducing instances based on conditions such as "Abandoned" and "Rule Infarction".

These preprocessing steps help prepare the dataset for further analysis and machine learning modeling.

Libraries-Used

Data-Pipelines-Preparation

Utilized machine learning algorithms to analyze results and select an efficient model for various data types. My approach includes the following steps:

• I used a DataFrameSelector class that inherits from BaseEstimator and TransformerMixin to select the required attributes of a data frame. This class has init, fit, and transform methods and accepts a list of attribute names.

• Two pipelines, num pipeline and cat pipeline, handle numerical and categorical data, respectively. The num pipeline selects numerical attributes and uses SimpleImputer to fill missing values with the mean. The cat pipeline selects categorical attributes and fills missing values with the most frequent value using SimpleImputer. It also applies LabelEncoder to encode categorical values into numerical values.

• The FeatureUnion function combines numerical and categorical pipelines into a single pipeline named data prep pipeline.

• The run model function takes a machine learning model, training, validation, and test data as input. It creates a full pipeline including the data prep pipeline and the given model. After fitting the pipeline on the training data, it returns the accuracy, AUC score, and other performance metrics of the model on the training, validation, and test data. The function also records the fit time of the model and adds all performance metrics to a pandas data frame named expLog.

Model-Processing

I have utilized a variety of algorithms to power its functionality. The decision algorithms employed include:

• Tree Classifier
• RandomForest Classifier
• Adaboost Classifier
• GradientBoost Classifier
• Cat-BoostClassifier
• XGBM Classifier
• KNN Classifier
• Gaussian Naive Bayes
• Multi-Layer Perceptron Neural Networks

For the Flask application files used:

• Model.py: Responsible for running user moves against the model and providing predictions.
• flask app.py: Handles user moves and provides results.
• chess engine.py: Provides the GUI for the chess interface.
• board test.py: Used for testing and validating user chess moves.

Instructions

1. CREATE AN .env file

python3 -m venv <virtual-environment-name>

2. ADD THE FOLLOWING LINES IN IT

FLASK_APP=flask_app.py
FLASK_DEBUG=1
FLASK_RUN_PORT = 3000
APP_SECRET_KEY = IOAJODJAD89ADYU9A78YGD

3. THEN RUN THE FOLLOWING IN THE TERMINAL WITHIN THE ENVIRONMENT

pip3 install -r requirements.txt

4. TO RUN THE SERVER USE

flask run

OR

python3 flask_app.py

5. GENERATE API KEY FOR GEMINI AI

To integrate Gemini AI, you need to obtain an API key from the Gemini AI platform. Insert this API key in your flask_app.py file to authenticate requests to the Gemini AI API.

Project-Demonstration

Shatranj-Demo.mp4