Lecture on Moving Birds Using Neural Networks

Introduction

• Goal: Move birds based on a neural network
• Previous steps: Checked collisions, pipes, and removals

Inputs for Neural Network

• Distance between top pipe and bottom pipe
• Y position of the bird
• Handling multiple pipes on the screen
  • Max two pipes at all times
  • Need to determine which pipe to focus on

Pipe Selection Logic

• Initialize pipe_IND = 0
• Condition to check which pipe to focus on:
  • If len(birds) > 0
  • If len(pipes) > 1
    • Compare bird’s X position with pipe’s X position and width
    • Adjust pipe_IND accordingly

Moving Birds

• For each bird:
  • Get neural network output
  • Move the bird: bird.move()
  • Add fitness for survival: bird.fitness += 0.1 each frame
  • Activate neural network: output = net.activate()
  • Inputs for activation:
    • bird.Y
    • abs(bird.Y - pipe[pipe_IND].height) (top pipe distance)
    • abs(bird.Y - pipe[pipe_IND].bottom) (bottom pipe distance)
  • Condition for bird to jump: if output[0] > 0.5 then bird.jump()

Additional Conditions

• Check if bird reaches the top of the screen (bird Y position < 0)
  • Bird should not survive if it does
• Correct loops and function calls (e.g., drawing birds, quitting game on no birds left)

Testing and Errors

• Birds performing well with random neural networks
  • Large number of birds (e.g., 100) covers many neural network combinations
• Reducing number of birds affects performance
  • Demonstrates need for neuroevolution

Training and Fitness Function

• Change fitness functions to observe effects
  • Minor changes greatly affect bird performance

Statistics and Generation Details

• Population's average fitness, standard deviation, best fitness, etc.
• Observations from running multiple generations
  • Quick improvement due to good fitness function

Saving and Using Best Bird

• Introduction to saving models with pickle
• Pickle the best genome and load it later to use its neural network
• Importance of fitness threshold and breaking process in training
  • Example: Break loop if score > 50

Conclusion

• Reminder to modify fitness functions and experiment
• Encouragement to subscribe to the channel for more tutorials

Final Notes

• The tutorial series demonstrates the integration of neural networks for game playing
• Audience encouraged to apply learnings and explore further