Ai

Motivated by my internal drive, I am deeply passionate about exploring AI systems in game development, with a focus on enhancing user experiences and streamlining production processes.

Pose Transfer - MSc Thesis Project

Recently I completed my master's in AI and Ml from Liverpool Jhon Mores University. I have taken Pose Transfer as my thesis topic. Here is the content of my video submission.

Quick Summery of Research

Data and it's distribution

Semantic map Genaration

In the data pre-processing stage, we faced the challenge of lacking segment maps in the dataset, which are important for our proposed method. • To address this challenge, we utilized a human parser proposed by Gong et al. to extract 12 distinct body parts and generate human segmentation maps

Network Architecture

Network Architecture

We have used two generators: The semantic parsing map generator and the Main generator. • We have used three encoders: Semantic map encoder, Heat map encoder and Source image encoder. • We have used an attention mechanism in every resolution to avoid data loss during the conversion of latent space. • We have created a Discriminator network to create adversarial training.

Training

• The network trained in two phases red and violet represents in the graph. • Loss of generator reached local minima and picked upward trend

Qualitative comparission

Qualitative comparission

Results

Area of Improvements

• The method utilizes attention links and a semantic parsing map in the generator's encoding and decoding streams, resulting in an improved network architecture. • The approach achieves state-of-the-art results on the challenging Deep Fashion dataset, competing with existing methods in both qualitative and quantitative benchmarks. • The proposed network is effective in various application domains, including virtual try-on, and data augmentation. • The end-to-end architecture provides a flexible and efficient framework for human pose transfer and related tasks. • The use of attention links enables selective and accurate pose transfer, while the modular design allows for easy adaptation to domain-specific purposes. • The research offers new possibilities for applications in virtual reality, gaming, and fashion

Win Rate Prediction for OZMM

Overview:
Win rate prediction involves iteratively tuning game levels to achieve a balance between actual and expected performance. This process often requires multiple iterations to stabilize win rates, making it both time-consuming and resource-intensive.

Problem Statement:
The goal is to predict the win rate of a level during its design phase, reducing the need for extensive manual iterations and optimizing the design process.

Proposed Approach:

Leverage the dataset of approximately 6,400 levels with their corresponding win rates.
Implement a machine learning model to predict win rates for new levels based on historical data.

Challenges:

The level data is stored in text format, which must be converted into a structured tabular format suitable for machine learning.
Mapping the transformed data to corresponding win rates requires careful preprocessing and feature extraction.

Example level file:

Text to Tabular format Conversion

I have created a script to convert this data into a tabular format, but the process is quite time-consuming.

Model Building

Approach 1: Linear Regression
I initially used a linear regression model to predict the win rate, but it was unsuccessful. Upon analysis, I realized that there is no linear relationship between the input variables (X) and the output variable (Y).

Approach 2: Tree-Based Models
Next, I experimented with various tree-based models, including Random Forest, Gradient Boosting, and XGBoost. While all models faced challenges, XGBoost outperformed the other two.

Evaluation Metrics:
To measure model performance, I utilized standard metrics such as Mean Squared Error (MSE) and R² Score.

Graph of actual winrate vs predicted winrate:

Approach 3: Deep Neural Network (DNN)

I implemented a Deep Neural Network (DNN) to predict the win rate. The model was fine-tuned using techniques such as dropout and normalization.

The training process lasted for approximately 22,000 epochs, taking about two days. However, the model failed to perform well, with a training loss of 50–55 and a validation loss of 150–160. This highlighted the need for more data to make the DNN approach effective.

To address this, I simulated additional data, increasing the dataset size from 6,400 to 32,000 samples.

After retraining the model, performance improved significantly, with a noticeable increase in the R² score.

However, the model still failed on the test data, indicating an overfitting issue.

Approach 3: Improvements

To tackle the overfitting issue, I incorporated batch normalization and additional dropout layers into the network architecture. Additionally, I transformed the win rate data into categorical bins and analyzed the distribution across these bins. This analysis revealed a significant class imbalance in the data.

I simulated the data while maintaining its original distribution and verified it using statistical measures.

This approach allowed me to achieve a balanced dataset with respect to win rate.

I also included two additional metrics: low win rate range and high win rate range into the data.

Result:

After training for about 500 epochs, the results during both training and validation were very promising:

I also tested the model with new data, and the performance was solid. The error was reduced to 36, and the R² score improved to 77%.

The result is very close compared with actual winrate.

Graph of Actual and predicted winrate

Next Steps:

I will develop a basic interface and provide it to the designers for testing on future levels. The results will then be compared with the actual player win rates.

Additionally, I plan to enhance the model by incorporating another metric: win rate bins. This can be predicted using a classification model.

Conclusion:

I took this project personally, dedicating my own time, weekends, and extra hours to complete it. Having completed my master's in AI and ML, I wanted to apply those skills to a real-world problem, which I hadn't had the chance to do before. This project has been one of the most challenging one, and I truly enjoyed the process.