Info
This post was imported from a personal note. It may contain inside jokes, streams of consciousness, errors, and other nonsense.
Last time I started off by trying to get a neural network running from an example online but I think I’m diving into the deep end without a life jacket. I was using this example: https://www.geeksforgeeks.org/ml-neural-network-implementation-in-c-from-scratch/
Anyway, I’m rolling it back to a very simple implementation with two inputs for the sensors, the extra constant node (what’s it called again?), and two outputs for the propellors. I’ll just use std::vectors of floats for now and figure out how to use Eigen for it later. Could use a review of that Andrew Ng course I took … at the end of 2020. Well a lot’s happened since then.
It’s called a “bias.”
Sweet. A simple neural network using two inputs and two outputs actually does the trick. Even one output of turn direction works but I added a neuron to control speed as well.
This is using hard-coded weights in the network for now.
weights({ -0.2, 1, 0, 0, 0, 1 })
The first value uses the bias to get the boids to slow down gradually over time. The second value accelerates when food is detected, more acceleration the closer the food is. The last value turns in the direction of the food.
I ended up not using explicit sensors or propellers for now. It’s just easier. I used two inputs: food detection (inverse of distance), food direction (angle between boid direction and food source). And I used two outputs: speed (accelerate or decelerate), turn.
I’m not sure it’s possible with this setup to turn toward food only when it’s detected. Might need more neurons for that or maybe go back to the Braitenberg vehicles model with spatially separated sensors.
A few things I need to do to “evolve” the weights instead of hard-code them. Evaluate the fitness of each boid then take the top boids and clone them with some mutations to the weights. That means I’d need to be able to restart the simulation with new brains so saving and loading files would be great.
What do I want this workflow to look like?
Surely I don’t want to create a whole GUI for this. I can do most of it on the command line and just run the current app to get a visualization of their behaviour. I’ll need another app for running the sim. So workflow would be:
- Use a script to create a file representing a specified number of boids with neural nets initialized with random weights.
- Use the command line to run many many iterations of the simulation with boids from the file and save their new fitness scores to file. (Should I save several properties and calculate the fitness score from them or just save the fitness score?)
- Use a script to retrieve the top boids and repopulate with mutated version of them. Output the result to a file. I might end up folding this into step 2 down the road so I can run multiple generations with a single command.
- Run the visualization with boids from the file. Or run the simulation on them again.
- Right now the simulation runs out of food sources. Should I automagically repopulate so the simulation can run forever? Try to place them where they’re not already colliding with a boid?
File Format #
Each boid needs its weights and a fitness score. I guess that’s really it. CSV would be more convenient.