Below is the reformatted Markdown document with improved structure, clarity, and consistent code blocks:

---

Stardew Valley Fishing Minigame Double DQN

Overview

This project implements a Deep Reinforcement Learning agent to master a fishing minigame in Stardew Valley. Using a Double Deep Q-Network (DDQN) architecture, the agent learns to time button presses perfectly to catch various fish with different movement behaviors and difficulty levels.

Technical Features

• Double DQN Implementation: Reduces Q-value overestimation bias for more accurate decision making.
• Advanced Network Architecture: Utilizes a 3-layer neural network (128-128-64) with optimized weight initialization.
• Curriculum Learning: Gradually introduces harder fish types as the agent improves.
• Time-Aware State Design: Incorporates time as a state dimension for better strategic decisions.
• CUDA Acceleration: GPU-optimized tensor operations for faster training.

Installation

Clone the repository and install the dependencies:

# Clone repository
git clone https://github.com/keethesh/StardewValleyFishingAI.git
cd StardewValleyFishingAI

# Install dependencies
pip install torch numpy matplotlib tkinter

Usage

Training a New Model

To train a new model, set train_new_model = True in the main block and run:

python main.py

Using a Pre-trained Model

To use a pre-trained model:

1. Set train_new_model = False in the main block.
2. Ensure the model path is correctly specified.
3. Run:

python main.py

Interactive Mode

After evaluation, the program enters interactive mode, where you can:

• Select specific fish to watch the agent catch.
• Choose random fish.
• Observe performance across different fish behaviors.

Fish Behaviors

The agent learns to handle five distinct fish behaviors:

• Mixed: Combination of movement patterns.
• Dart: Quick, sudden movements requiring fast reactions.
• Smooth: Gradual, predictable movements.
• Sinker: Downward-biased movement.
• Floater: Upward-biased movement.

Training Process

1. The agent starts with simpler fish (difficulty ≤ 40).
2. As performance improves, it progresses to medium difficulty (≤ 70).
3. Finally, it tackles the hardest fish (≤ 100).
4. Early stopping occurs after a consistent 98%+ success rate.

Model Details

• State Space: 11 dimensions, including positions, speeds, fish characteristics, and time.
• Action Space: 2 actions (press/release button).
• Network Architecture: 3 hidden layers with sizes 128, 128, and 64 neurons respectively, using ReLU activations.
• Learning Algorithm: Double DQN with soft target updates.

Best Models

Pre-trained models are saved in the "Best models" folder from multiple training runs. Each model represents a different training session with various hyperparameters and optimizations.

Results

• Overall Success Rate: 95% across all fish types.
• Smooth-Movement Fish: 92.7% success.
• Dart-Movement Fish: 78.4% success.
• Higher Difficulty Fish: Near 100% success.

Acknowledgments

This project demonstrates the application of recent reinforcement learning techniques to a timing-based game, showcasing how curriculum learning and network architecture choices can significantly impact learning efficiency.

---