Optimising outcomes by gradient descent

[calumroy]

The Botech protocol is a cool idea. It made me think about adding the ability to make a differential graph using the prototcol. Then you could apply standard gradient descent to apply optimisation techniques.

E.g maybe you want to answer the question how can I keep the number of people killed from cancer below a certain amount whilie minimising the amount of money initailly spent on treatment. The number of people killed is an output on the graph and the money spent an input (maybe that's not how the graphs work?). Apply gradient descent by finding the error difference between the last simulated number of deaths output and the desired number then update each node and propagate the errors back through the graph edges. Use some stadard gradient descent like ADAM or something to step the graph through updates to obtain an eventual required inital investment.

There probably plenty of other uses for this type of graph.

[RoryWatts]

That's really interesting Cal, thanks.

The number of people killed is an output on the graph and the money spent an input (maybe that's not how the graphs work?).

Yeah the graphs can definitely be set up this way.

Use some stadard gradient descent like ADAM or something to step the graph through updates to obtain an eventual required inital investment.

Have you ever used Optuna? Not that its necessarily equivalent, but it's been a cool package to use with gradient descent capabilities.