SFML community forums
General => SFML projects => Topic started by: lolz123 on September 24, 2014, 04:54:14 am
-
Hello!
As some of you may know, I am really into AI, especially reinforcement learning.
I use SFML to set up experimentation environments and to visualize them.
So, I thought I should post here about what is now a rather large collection of reinforcement learning agents.
I am still using the same cart-pole balancing simulation and mountain car problems, but I recently also added a maze.
Perhaps some of you find it useful for a game or research!
Link to repository: https://github.com/222464/AILib (https://github.com/222464/AILib)
Here is the list of available agents and components of the agents, as of 9/23/2014:
- APSSC - ant colony optimization based labeler
- Evolved CTRNN - a second-order recursive neural network used for evolution
- ConvNet - 2D convolutional neural network for feature extraction using restricted boltzmann machiness
- Autoencoder - unsupervised learning and feature compression
- DBN - deep belief network (stacked RBM)
- DSOM - deep self-organizing map for massive feature reduction
- FA- function approximator (standard multi-layer perceptron)
- FERL - free-energy based reinforcement learner. Simplest learner in this library
- RBM - generic restricted boltzmann machine
- DNF - dynamic neural field reinforcement learner (hebbian based)
- Elman - simple Elman network implementation
- TDFalcon - ART (adaptive resonance theory) based reinforcement learning
- HTM - hierarchical temporal memory
- HTMRL - continuous action and state POMDP reinforcement learner
- HTMRLDiscreteAction - discrete action and continuous state POMDP reinforcement learner
- Bayesian Optimizer - simple implementation
- HyperNet - proto-ERL
- LSTMG - generalized LSTM implementation
- LSTMActorCritic - actor-critic using LSTMG
- LSTMRL- simpler version of LSTM discrete action reinforcement learner
- ActorCriticAgent - backprop critic, hebbian actor
- CACLA - continuous actor-critic learning automaton
- MemoryActor - actor portion only with hebbian learning and memory cells
- MultiQ - simple Q learner with function approximation and discrete actions
- NCPSOAgent - no critic particle swarm optimization agent
- PSOAgent - critic included particle swarm optimization agent
- QAgent - continuous state and action Q learner using iterative policy derivation
- RLLSTMAgent - simplified LSTM continuous action and state reinforcement learner
- SOM - simple Kohonen self-organizing map
- SOMQAgent - SOM-based state discretizing Q learner
- TabularQ - simplest possible discrete state and action Q learner
- RAAHN - real-time autoencoder augmented hebbian network
-
This is a huge, great and super interesting collection about AI!
Thank you for sharing your work, this will definitely be helpful.
-
This does look interesting, however I highly suggest you clean up your git repository. Take out all of the temporary files, binaries, sfml libs, and anything else that really doesn't specifically belong to the project. It would also be nice if you could add a cross platform build system to your repo so it can easily be tried out without having VS installed. ;)
-
@zsbzsb: Yes I know, I got lazy :P . Unlike my ERL project (where I set up a cross-platform build system and all that properly), it isn't really intended to build on its own; rather, it makes more sense to pick out the agent you want to use/modify and modify/include the source in your own project for that agent. It's not really a library, more of a reference ;).
-
I've fixed a few non-standard conforming things (std::sinf std::fmodf etc.), cleaned up the repository and added CMake support.
Check your pull requests: https://github.com/222464/AILib/pulls
Now I can start looking at the actual code and stuff. :D
-
@exploiter: Wow! Thank you so much! I will merge it as soon as possible! I didn't think I would see so much interest in this project ;D
By the way, I made a Reddit post on my ambitions with a particular agent in this repository, HTMRL: http://www.reddit.com/r/MachineLearning/comments/2hdc59/my_attempt_at_outperforming_deepminds_atari/ (http://www.reddit.com/r/MachineLearning/comments/2hdc59/my_attempt_at_outperforming_deepminds_atari/)
It is probably the most powerful agent in the repository right now, but also the most complicated!
-
There's normally a lot of interest in making AI's but most are looking for how to make them rather than for someone like you who has saved them the trouble of doing so.
Yes I know I'm necroing. :(
-
This project still active? Was planning on seeing how much of the different AIs I could work on changing over to C#.
-
This project still active? Was planning on seeing how much of the different AIs I could work on changing over to C#.
This project (AILib) is no longer active - however, there is a new one I am working on for a company called Ogma.
The name of the project is called OgmaNeo. It's a versatile, fully online hierarchical learning algorithm that runs on the GPU (or CPU if desired). It actually does still use SFML for the demos, though the library itself only depends on Flatbuffers and OpenCL.
Link to the repository: https://github.com/ogmacorp/OgmaNeo (https://github.com/ogmacorp/OgmaNeo)
Due to its online nature, it is also super fast, and may actually be useful for game development (what these forums seem to focus on). Indeed, one of the demos is a level generator that uses pixels to create infinite levels.