Decision Transformer for Game Input Prediction

This project implements a Decision Transformer model for predicting game inputs based on state observations. It's designed to learn from gameplay data and generate mouse and keyboard actions.

Setup

1. Clone the repository:

   git clone https://github.com/yourusername/decision-transformer.git
   cd decision-transformer

2. Create and activate a conda environment:

   conda create -n dt-env python=3.8
   conda activate dt-env

3. Install the required packages:

   pip install -e .

Data Preparation

Convert your CSV gameplay data to SafeTensor format using the provided script:

python scripts/preprocess_data.py --input_file path/to/your/csv/file.csv --output_file path/to/output/safetensors/file.safetensors

Configuration

Adjust the hyperparameters and settings in configs/config.yaml:

Training

To train the model, run:

python scripts/training.py

Model Architecture

The Decision Transformer is based on GPT-2 and processes state observations, mouse actions, and key actions to predict future actions. Key components:

1. DecisionTransformer: Main model class
2. GPT2Model: Underlying transformer architecture
3. Trainer: Handles the training loop and evaluation

Customization

• Modify decision_transformer/models/decision_transformer.py to adjust the model architecture.
• Update decision_transformer/training/trainer.py to change the training process.
• Alter data/data_loader.py to modify how data is loaded and processed.