Samuel Brown
Software developer in Wokingham, UK. I build small models of things I'm curious about, sheepdogs and crop diseases so far, and write up what they taught me.
All postsThe robot collie
The same sheepdog problem, tackled three ways. It starts with two rules from a paper that turn out to pen a flock, and I keep adding awkward sheep and obstacles until they stop working. Then the rules go away and a small neural network has to work the job out for itself, evolved one litter at a time, and the last post asks whether you can tell early which training runs to keep.
- The collie is the algorithm
Someone worked out that you can simulate a sheepdog with two rules. I didn't believe it, so I built it, then kept adding awkward sheep and obstacles to see when it would break. There are four simulations in here you can play against.
- Making a robot collie
Last time I wrote the sheepdog's rules by hand. This time the dog had to work them out for itself, one generation at a time. You can evolve one in your browser, race it, and watch what it's thinking while it runs.
- Picking the puppy
Evolving the robot collie worked, but some of the runs only worked because they got lucky. So I trained two hundred and twenty of them, three different ways, to find out whether you can tell early on which runs are going to come good. There's a game in here where you can have a go at picking one yourself.
Forecasts for farmers
Britain issues warnings to farmers that rest on rules written decades ago and tested, if at all, only against the years when the disease turned up. I rebuild each one from the same public weather and outbreak records, score it against the days when nothing followed too, and try to find out what the warning is actually telling you.