FlappyBirdAI

Performance Metrics & Accuracy

In Reinforcement Learning (RL), “accuracy” isn’t measured like in classification tasks (e.g., “99% correct”). Instead, we measure performance by the Total Reward accumulated per episode and the Average Score over time.

The Reward Function

The agent learns by trying to maximize its cumulative reward. We define specific rewards and penalties to guide its behavior:

Event	Reward	Reasoning
Stay Alive	+0.1 (per frame)	Encourages the bird to survive as long as possible.
Pass Pipe	+1.0 (implicit)	Accumulating “Stay Alive” rewards while passing pipes results in a high score.
Flap (Jump)	-0.2	Penalizes excessive flapping to encourage efficient movement and prevent “jittering” at the top of the screen.
Death	-100	A massive penalty for hitting a pipe, ground, or ceiling. This is the most critical signal for the AI to avoid.

Training Progress

Epsilon-Greedy Strategy

The agent uses an Epsilon-Greedy strategy to balance exploration and exploitation.

• Exploration: Taking random actions to discover new possibilities.
• Exploitation: Using the trained neural network to make the best decision.

graph TD
    A[Start Episode] --> B{Random < Epsilon?}
    B -- Yes --> C[Random Action]
    B -- No --> D[Model Prediction]
    C --> E[Execute Action]
    D --> E
    E --> F[Observe Reward]
    F --> G[Update Epsilon]

• Initial Epsilon: 1.0 (100% random)
• Decay Rate: 0.995 per episode
• Minimum Epsilon: 0.01 (1% random)

As training progresses, epsilon decreases, meaning the agent relies more on what it has learned and less on random guesses.

Measuring Success

We consider the model “successful” when:

1. Average Score consistently increases.
2. Epsilon has decayed to its minimum (0.01).
3. The bird can survive indefinitely (or for a very long time) without crashing.

This site is open source. Improve this page.