Exploring Fuzzy Logic for AI Decision-Making

🖥️ View the GitHub Repository

🎮 Check out my other projects

Video Demonstrations

The following videos showcase the presentation of this project and a demonstration of how fuzzy logic performs under different rule sets.

Presentation Video

Fuzzy Logic AI Test Results

Introduction

For this project, I implemented fuzzy logic AI in a Unity-based simulation, where a green car avoids obstacles (red cars) and collects items (purple coins).

Fuzzy logic is often underutilized in game development but has been seen in RTS games for opposition reaction systems. The goal of this coursework was to evaluate how “accurate” fuzzy logic can be in real-time AI decision-making.

Why Fuzzy Logic?

Fuzzy logic operates on a scale between 0 (false) and 1 (true) rather than strict binary decisions. This allows for more nuanced AI behavior by assessing inputs as degrees of truth rather than fixed values.

I was particularly interested in:

• How “fuzzy” decisions impact AI movement.
• How adding more rules affects AI accuracy.
• How well fuzzy logic can handle dynamic decision-making in games.

Implementation & AI Behavior

The fuzzy logic system was integrated into a green car game object, controlling its movement based on the positions of red cars (obstacles) and purple coins (collectibles).

Inputs Used in the System:

1. redCarDirection – Determines if the green car is left, right, or centered relative to an obstacle.
2. redCarApproach – Measures how far or close the car is to an obstacle.
3. itemDirection – Determines if the green car is left, right, or centered relative to a collectible.
4. itemApproach – Measures how far or close the car is to a collectible.

These inputs were processed using the Fuzzy Logic Library provided in the CMP304 Unity Lab.

Defining AI Rules & Testing Scenarios

To test how fuzzy logic affects decision-making, I created multiple rule sets:

Phase One – Arbitrary Values (Baseline)

• Initial logic rules applied without optimization.
• AI behavior was unpredictable and inconsistent.

Phase Two – Improved Input Values

• Adjusted input peaks to better distinguish object distances.
• AI exhibited more controlled movement but still had inconsistencies.

Phase Three – Improved Output Values

• Optimized defuzzification for smoother AI transitions.
• AI improved in avoiding obstacles and targeting collectibles more efficiently.

Testing Different Rule Sets:

• 14 Rules: Limited decision-making capability.
• 120 Rules: More refined decision-making, but with diminishing returns.

Key Takeaway: More rules did not always lead to better performance in this application.

Results & Observations

• Fuzzy logic was precise but not always effective.
• AI performed better when objects were farther away, giving it more time to react.
• More rules didn’t necessarily improve decision-making — in some cases, it overcomplicated the process.
• The best configuration was Phase Two, where input peaks were adjusted to reduce unnecessary collisions.

What Can We Learn from This?

• Adjusting input peaks directly affects reaction speed.
• Too many rules can reduce efficiency instead of improving accuracy.
• Fuzzy logic works well for AI movement but isn’t always the best choice for fast-paced reaction-based decisions.

Final Thoughts

This project provided valuable insights into AI decision-making and how fuzzy logic can be used in game development. While it has strengths in nuanced behavior, it also requires careful tuning to avoid overcomplicating AI logic.

This was an exciting experiment, and I look forward to further exploring AI-driven decision-making in games!