TensorFlow, AI, Autopilot, VendeeGlobe

[mgrouch]

I would like to suggest some detail on the first high level idea.

1)  Use learning networks to steer the boat as an alternative to basic filters.

• More nuanced polar performance predictions based on currently observed conditions, learned from previously observed similar conditions
  

On this one, I'm thinking about Sail selection, wave height and direction, and the obvious fact that increasing wind speed doesn't translate to increasing speed after a point. 

• Wave Pattern Recognition: Use sensors to recognize wave patterns and adjust boat movement accordingly.
  
• Learned optimal helm response with respect to observed conditions, including wave direction and height
  
• Current and Tide Optimization: Adjust route and speed based on tidal and current data.
  
• Adaptive Tacking and Jibing: Use ML to determine optimal tacking/jibing points based on current and predicted conditions.
  
• Dynamic Waypoint Adjustment: Adjust waypoints in real-time based on changing weather conditions and sea state.
  
• True Wind Angle Optimization: Adjust sailing angle dynamically for optimal speed and safety.
  

2)  Detect hazards to navigation and thus perform watch keeping ability.  Surely a computer can do a better job of this than a human.

3)  Literally steer around objects as well, but this is the most difficult to implement and also not all that useful compared to the above.  In any case,  #2 must be implemented before we have to worry about #3

Here are some of my ideas:

1. Add particle swarm optimization to auto tune autopilot gains

2. Add fractional order PID controller to autopilot FOPID with auto tuning 

https://www.mdpi.com/2076-3417/12/6/3139

3. For sea state detection and wave height and frequency, something like this
can be used: (Kalman filter with initial parameters from trochoidal wave model)
with Doppler effect formulas for moving boat

https://bareboat-necessities.github.io/m...-math.html

Simple approximation for wave height can be taken from
max acceleration and wave frequency in trochoidal model. 

Moreover vessel RAO (response amplitude operator) has to be taken into account. 
RAO is a complex number function of wave frequency and amplitude. 
RAO is like a dynamic fingerprint of a hull design and also load (like in cargo weight) distribution 
There will be a different RAO for each degree of freedom of the hull – the linear motions surge, sway, heave, and the orientation motions roll, pitch, and yaw.
If phase is not in consideration absolute value of complex value of RAO is used

4. Wave direction. How to properly detect it if it’s not aligned with wind?

https://www.mdpi.com/1424-8220/22/1/78/htm

5. Use Julia programming language instead of python. 

6. For object detection use yolov8 

https://github.com/ultralytics/ultralytics

YOLO (you only look once) uses CNN convolutional neural networks to process images fast

Fair winds!