Crate for implementing NeuroEvolution of Augmenting Topologies
License: MIT License
Because it's being constantly run after predictions anyway, it's probably wise to just merge the two and prevent the whole host of issues associated with it. Still not sure whether it's worth doing a refactor like this so late in development though.
The recursiveness of the is_connection_cyclic makes it hard for contributors to understand and should be refactored to a loop for readability.
Not sure how this is happening but in extremely rare circumstances it is possible to hang indefinitely. See #57 workflow run for more info.
My guess is there is one very small outlying situation that causes a rwlock to be locked and used by a child node, but this shouldn't be possible with the well-tested circulation prevention algorithm. This definitely requires further debugging, but it is so rare and obscure that it is difficult to catch it and the details about what happened.
Instead of using cfg on the entire main function, it's probably possible to get away with just doing certain lines from the main function
The example code does not work, you should update it to the one listed in examples folder.
Update just released. Doesn't change much internally but examples could use it.
There clearly aren't enough tests here, as I keep coming across runtime errors in production that should have been caught by the actions.
The original idea with having these split off was to have the NeuralNetwork struct inputs have smart pointers to the layer before, but this never really was implemented (and is likely impossible to implement without big performance costs). Right now it is just a bunch of useless cloning to access certain functions.
NeuralNetworkTopology uses Arc<RwLock<NeruonTopology>> but never uses async code. It would be better to only use something like Rc<RefCell<NeuronTopology>>.
There should be an option to perform crossover mutation by merging/averaging neurons instead of 50% chance of having a neuron from one network or the other.
Will likely implement the same way as mutation_rate and mutation_passes.
Just looks weird on the docs having all the sigmoid and ReLU and stuff with the rest of the activation registry stuff, should probably make a separate impl module or something to make it less confusing.
should use #[doc(feature = "something")] and #[cfg(any(feature = "something", doc))] to include feature documentation.
The NNTSerde struct makes it easier to write because of the derive macro, but it is also a lot slower than just converting directly from the neural network topology.
The topology.rs file is starting to get really long, so it would be good to turn this into a directory module soon to keep it maintainable and readable.
Allows AIs to run on the GPU for increased parallelization. Might be difficult/impossible to implement with all these RwLocks, probably requires a major rewrite.
By moving the cache to a separate HashMap or something, neurons would not need to be stored in a RwLock, greatly reducing the amount of data being cloned while transferring between topology and mutable. Additionally, flush_state could be merged into predict without fear of performance issues because the map can simply be replaced by a new one in memory, which is way faster than the current iterative approach.
Essentially adds it to the neuron topology/gene so that they can be more diversely calculated
Should rewrite or refactor some of the code to make it faster/more readable.
forgot that line, may reopen issue about flushing state just because I'm seeing how problematic it could be if forgotten.
Currently, nets are initialized by linking all output nodes with all inputs and applying sigmoid to all outputs. Instead of this, maybe could introduce some sort of "network constructor" trait that lets certain structs init a neural network (ex: SigmoidMLP::new<I, O>() -> NeuralNetworkTopology<I, O> or something of the sort).
Because this crate was sort of rushed, it is more of an abstract and unfamiliar implementation of NEAT rather than the actual algorithm. This should be solved later as this crate is further developed.
A trait or something to simply say "value at this index is this variant of enum" and then impl From<[f32; LEN]> automatically. Can definitely make it better with a derive macro.
Should add ability to serialize topologies/genomes
Following a comment I made in #50, the examples' fitness graphs look like this:
This does not have any sort of upward trend at all, just looks like random data.
First change could be to try running these examples with a lot of generations to see if it is just statistical noise from not having enough generations.
If that is not the case, it is likely either an issue with the fitness function or a much larger core issue that needs solving.
This will help introduce people to the new activation registry API and overall just enhance the ease of use of this crate.
Current add neuron mutation should use weight 1 for first connection and preserve weight for second.
Also should have remove neuron mutation.
I published a new update for that so this needs to be updated too.
While the most widely-used activation functions are predefined, it'd be better to have some sort of registry so that the serde deserialization works with custom activations.
It would also be good to note that there should be a separate flag or something for those activation functions which do not want to be randomly mutated to/fro.
Somehow (in very rare circumstances) the partial_cmp will return None. I believe this may be caused by #53.
thread 'main' panicked at /workspaces/neat/src/runnable.rs:296:57:
called `Option::unwrap()` on a `None` value
note: run with `RUST_BACKTRACE=1` environment variable to display a backtrace
Aborted (core dumped)
Possibly the cause of #58, as I noticed this in the neuron dump by @Bowarc (also in #58). A neuron can have multiple references to the same neuron as input. This, when combined with rayon, makes it so that, as it performs parallelized iteration on the neurons, there is a chance they try to write to the same neuron at the same time, causing a deadlock.
Should probably either change how I formatted the links or make it a markdown file.
I'm guessing this is somehow caused by sigmoid but I don't know how. check activation example branch for more info.
React
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
Typescript
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
Django
The Web framework for perfectionists with deadlines.
Laravel
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
Facebook
We are working to build community through open source technology. NB: members must have two-factor auth.
Microsoft
Open source projects and samples from Microsoft.
Google
Google ❤️ Open Source for everyone.
Alibaba
Alibaba Open Source for everyone
D3
Data-Driven Documents codes.
Tencent
China tencent open source team.