engine_release_changes.txt

来自「这是一个神经网络的开发工具」· 文本 代码 · 共 137 行

1.2.0
New Feature: Added a basic implementation of the RBF networks
New Feature: Added the BiasedLinearLayer, useful to implement the RBF networks
New Feature: Added the DeltaNormPlugin in order to pre-elaborate time series
1.1.1
Bug Fix: Fixed the 'No matching patterns' problem (bug #1024973)
Bug Fix: Fixed the globalError set to 0.0 for a cloned neural network (bug #1013956)
Bug Fix: Fixed the missing use of the momentum parameter in the BatchLearner (bug #1024960)
Bug Fix: Now the KohonenSynapse registers itself correctly as NeuralNetListener when deserialized (bug #1025177)
1.1.0
New Feature: Added the 'param' property to the Monitor object in order to permit to store whatever new parameter without the need to change the java code
New Feature: Added the RPROP algorithm (by Boris Jansen)
New Feature: Added the method join to the NeuralNet object in order to implement a mechanism to wait for all the running thread to finish
Enhancement: Revisited the threads synchronization to improve the training speed (improved more than 35%!)
Enhancement: Revisited the learners handling mechanism in order to be able to add/choose dynamically the Learner components
Bug Fix: Fixed a bug in batchmode (bug #991335) (by Boris Jansen)
Bug Fix: Fixed a wrong calculus of the output neighbourhood radius of the GaussianLayer component when in recall mode
Bug Fix: Fixed a bug in Layer.isRunning method (bug #1007838) (by Jerry R. Vos)
1.0.1
Bug Fix: Fixed a wrong initialization of the parent Monitor within the NestedNeuralLayer class
Bug Fix: Added the netStoppedError notification to the internal NN of the NestedNeuralLayer class
Bug Fix: The NestedNeuralLayer's randomize/addNoise method calls now have effect only when in learning mode
Bug Fix: Resolved a wrong calculus of the validation error after consecutive runs (bug #957664)
Bug Fix: Fixed the wrong validation error calculus when preLearning > 0 (bug #949353)
Enhancement: Eliminated an useless warning raised by the TeacherSynapse when stopped
1.0.0 final
New Feature: Added the setComponent method to the Inspection interface and its inherited classes
New Feature: Added the maxBufSize property to the input synapses based on the StreamInputTokenizer
Enhancement: Decoupled the engine from the logging libraries to be independent from the Log4j library
Bug Fix: The StreamInputSynapse checks now if its input patterns number is lesser than the Monitor's input patterns setting
Bug Fix: The FileInputSynapse opens the input stream only when requested
Bug Fix: Resolved the IOException 'too many open files' on FileOutputSynapse object
Bug Fix: Changed the tokenizer handling into all the classes inheriting StreamInputSynapse
Bug Fix: The NestedNeuralLayer now handles correctly the attached I/O components
Bug Fix: The XLSInputSynapse now handles correctly sheets containing empty columns/rows
Bug Fix: The XLSInputSynapse now handles correctly the sheet name property
Bug Fix: The Monitor object now handles correctly the nested Monitor, if present
1.0.0 rc1
New Feature: Added the ComparingSynapse to compare two separate output patterns
New Feature: Added the KohonenSynapse, WTALayer and GaussianLayer to build unsupervised SOM networks
New Feature: Added the batch learning mode
Enhancement: Revisited the inspecting mechanism eliminating any depends from the editor's packages
Enhancement: Added a notification mechanism to advise all the running thread in case of a failure occurs
Bug Fix: The File/XLSOutputSynapse objects appended the patterns at the end of the output stream. Fixed by adding the property 'append'.
Bug Fix: The Layer.removeAllInputs/Outputs methods didn't reset the corresponding Synapse's input/outputFull properties
0.9.9
New Feature: Added the SnapShotPlugin recorder to create copies of a neural network at predefined intervals of time during its training phase
New Feature: Added the YahooFinanceInputSynapse to read stock exchange time series from the Yahoo's internet service
New Feature: Added the JDBCInputSynapse to read patterns from a wide range of database systems such as ODBC,Oracle,MySQL , Sybase as long as you have the database Vendors JDBC Driver
New Feature: Added the UnnormalizerPlugin class to de-normalize the output of a neural network
New Feature: Added the OutputConverterPlugin abstract class to implement any kind of post-processing plugin
New Feature: Added the SangerSynapse to build PCA - Principal Component Analisys - neural networks
New Feature: Added the MovingAveragePlugIn class to calculate the MA on a time series
New Feature: Added a NetChecker object that, along with the new check() method added to all the components, permits to check the correctness of a neural network
New Feature: Added an inspection mechanism to extract all the internal values of a component for analisys purpose
New Feature: Added an advanced serie selector property to the ConverterPlugin class to select as input several not contiguous columns
New Feature: Added an advanced column selector property to the StreamInputSynapse class to select as input several not contiguous columns
New Feature: Added a 'parent' property to the Monitor object to make possible the independency of the learning state of the NestedNeuralLayer from the parent neural network
Enhancement: Added a control mechanism that prevents a synapse to be attached simultaneously to different layers
Enhancement: In the Monitor object changed the name of the 'patterns' property to 'trainingPatterns'
Enhancement: Moved the validationPatterns property from the LearningSwitch to the Monitor object
Enhancement: Substituted all the System.out/err calls with the Log4J logging mechanism
Enhancement: To avoid that the org.joone.engine package depends on the org.joone.io package, the InputSwitchSynapse class has been moved to the org.joone.io package
Enhancement: To avoid that the org.joone.engine.learning package depends on the org.joone.io package, the LearningSwitch class has been moved to the org.joone.util package
Enhancement: Updated all the imports of the POI packages to adapt them to the new jakarta-poi 1.5.x
Enhancement: To avoid race conditions, a new class has been created that raises the netStopped event from a separate Thread  
Bug Fix: Fixed a race condition caused by the wrong management of the InterruptedException raised after a Thread.interrupt call on a running Layer
Bug Fix: Fixed a NaN output value on the NormalizerPlugin when all the values of a serie are the same
Bug Fix: Fixed a NullPointerException when the TeachingSynapse.setMonitor method was called with a null parameter
Bug Fix: Now the teacher is disabled if the network runs in recall phase, as in this case it doesn't matter anymore if a desired file is attached or not
Buf Fix: Fixed an ArrayIndexOutOfBoundsException raised when a net containing a DelaySynapse component was run
0.9.8
New Feature: Added the management of the macros into the NeuralNet object. Now all the BeanShell scripts are contained and managed (also transported) by the NeuralNet object
New Feature: Added the cloneNet() method on the NeuralNet object to create a copy of a neural network
New Feature: Added the NotSerialize interface to declare not serializable some objects during the exporting phase. Valid only for the classes inherited from the InputPatternListener and OutputPatternListener interfaces. 
New Feature: Added the boolean Exporting property to the NeuralNet Object, useful in conjunction with the NotSerialize interface. When it's true, all the classes implementing the NotSerialize interface aren't serialized.
New Feature: Added a mechanism to store/retrieve parameters to/from the NeuralNet object
New feature: Added the LogarithmicLayer component
New feature: Added the ContextLayer component
Enhancement: Added the SynapseBeanInfo class to permit to choose which properties are visible in the editor
Enhancement: Modified some internal mechanism to enhance the running speed
Bug Fix: Deleted an useless error message when invoked the Layer.stop() method
Bug Fix: TeacherSynapse calculates now the RMSE (root mean squared error) and not more the GSE (global squared error), making easier to compare the performances of the neural networks from external frameworks
Bug Fix: Added the resetInput() method also to the TeacherSynapse and TeachingSynapse objects
Bug Fix: Added the management of the 'enabled' property also to the xxxOutputStream and TeacherSynapse objects
Bug Fix: NestedNeuralLayer maintains now the connections made before the NestedNeuralNet property is set. In this case it acts like a LinearLayer component
Bug Fix: Layer now recognizes when it can't start because has not input synapses connected. It throws a RunTimeException now giving a proper message
Bug Fix: NeuralNet recognizes when on the start() method some layers can't start. It traps the error now and resets the net
Bug Fix: Changed the management of the i/o patterns in the synapses to make possible to build recurrent neural networks
0.9.7
Enhancement: The buffer size of the FileInputTokenizer changed from 64Kb to 1Mb
Examples Enhancement: Added a new example to show the use of a trained neural network
Bug Fix: Fixed the synchronization of the start of a neural net respect the end of the previous cycles
0.9.6
New Feature: Added the ScriptingPlugin to manage the net's events from a script contained in a string or file
New Feature: Added the getLayers to the NeuralNet object to extract the vector containing all the layers of the net
New Feature: Added the NeuralNetLoader class to load a serialized neural network
New Feature: Added the errorChanged event to the NeuralNetListener interface, and modified all the connected classes to manage the new event
New Feature: Added the 'validation' parameter to the Monitor object to indicate the validation phase
New Feature: Added the NeuralNetValidator object to test a neural network on a validation data set
New Feature: Added the LearningSwitch synapse to automate the switching of the input data set between the training and the validation set depending on the validation parameter of the Monitor object
Bug Fix: Fixed some bugs on the StreamInputSynapse to avoid an error when an input synapse has no layers attached
Bug Fix: Now the XLSInputSynapse can read not numeric cells
Bug Fix: Resolved some problems about the serialization of the XLSInputSynapse object
0.9.5
New Feature: Added the InputSwitchSynapse object to permit to handle many inputs
New Feature: Added the OutputSwitchSynapse object to permit to handle many outputs
New Feature: Added the NestedNeuralNetwork object to built nested neural networks
New Feature: Added clearest labels to some properties modifying the beaninfo classes
New Feature: Added in the org.joone.io the Excel I/O components
New Feature: Created the NeuralNetRunner class to run a neural net outside the GUI editor
New Feature: Added the 'getLayer' method to the NeuralNet object to retrieve a Layer by name
Enanchement: Refactoring of the org.joone.io's object model to permit to add new input data type
0.9.0
New Feature: Added the methods to get the actual engine version in the NeuralNet object
New Feature: Added the method randomize to the Layer and Synapse objects
New Feature: Added the method resetInput to StreamInputSynapse to reset the input buffer and read again the input stream
Bug fix: Changed the mechanism to stop the neural net enhancing its reliability
Bug fix: Changed the mechanism to restart a neural net
Usability enhancement: Changed a property name of org.joone.util.MinMaxExtractor to render it more undenstandable
0.8.6
Example enhanchement: Created a samples directory. The XOR example has been moved there.
0.8.5
Bug Fix: Now the CPU usage doesn't go to 100% before running the net
Bug Fix: Reingeneering of some components to make the engine more stable
0.8.4
Example enhanchement: Added the xor input file name on the command line to xor.java 
0.8.3
Bug fix: Setting the fileName before firstCol/lastCol, StreamInputSynapse object reads only the first row
Bug fix: Setting the fileName = "" throws a FileNotFoundException in StreamInputSynapse object
0.8.2
Automated the Monitor object passing from Layer to Synapse objects
0.8.1
Bug Fixes: the methods Layer.setMonitor sets now also the monitors of all input synapses
0.8.0
First release on SourceForge.