 Joshua Bowren
|
5e5f3b60b4
First public release.
|
8 år sedan |
 Joshua Bowren
|
93a9d4870a
Added online error reporting via the network visualizer. Split noise XML parameter into output and weight noise. Changed modulation to zero when user control is used. Changed learning rate of the Hebbian trained connection group in DefaultAutoencoder.xml to 1.0.
|
8 år sedan |
|
Joshua Bowren
|
3b43fa29c3
Started using libraahn with reconfigure and saturate. Fixed modulation scheme to not modulate against a wall when the last wall is not the same as the current. Switched learning rates back to 0.1.
|
8 år sedan |
|
Joshua Bowren
|
7643a851c7
Added training based off of a history buffer for the autoencoder component. Fixed a typo causing Hebbian training to always be used. Removed input scaling in libraahn version. Added frames around GL widgets. --experiment parameter checks Experiment folder after checking the local path. Added modulation debug display in the network visualizer. Changed the network files to use a learning rate of 5.0 and modulation feeler length of 700. Changed default sensor file to have a 180 degree angle between the first and last range finder. Also made ranger finders slightly shorter. Pure Hebbian learns to navigate the track consistently. RAAHN learns to navigate the track, but sometimes changes direction.
|
8 år sedan |
|
Joshua Bowren
|
87c4d489cf
Added a visualizer for neural networks. Changed the default experiment to only use range finders and one input group for them. Added DefaultAutoencoder which is similar to default, but adds a hidden layer whose incoming weights are trained using autoencoder training. Weights from the hidden layer to the output layer are trained with the Hebbian training. Added sensor file as an example for adding pie slice sensors.
|
9 år sedan |
|
Joshua Bowren
|
32fc9b9c74
Added XML based neural network creation. Added modulation and control schemes via ModulationScheme and ControlScheme. Added a modulation scheme for wall avoidance. Added a control scheme for turning based on sensor input. The current configuration has the agent set up for avoid walls. Pie slice sensor neuron connections are not modulated. Can be configured via XML.
|
9 år sedan |
