Changelog

All notable changes to this project will be documented in this file. The format is based on Keep a Changelog.

[0.18.0-SNAPSHOT] - Unreleased

Full changelog

Added

LearnLib now supports JPMS modules. All artifacts now provide a module-info descriptor except of the distribution artifacts (for Maven-less environments) which only provide an Automatic-Module-Name due to non-modular dependencies. Note that while this is a Java 9+ feature, LearnLib still supports Java 8 byte code for the remaining class files.
Added an InterningMembershipOracle (including refinements) to the learnlib-cache artifact that interns query responses to reduce memory consumption of large data structures. This exports the internal concepts of the DHC learner (which no longer interns query responses automatically).
The ADTLearner has been refactored to no longer use the (now-removed) SymbolQueryOracle but a new AdaptiveMembershipOracle instead which supports answering queries in parallel (thanks to Leon Vitorovic).

Changed

The JPMS support introduces several changes:
- You now require at least a JDK 11 to build LearnLib.
- We use modules to better structure the aggregated JavaDoc. Since there exist breaking changes between Java 8 and Java 9 regarding documentation (see package-list vs. element-list), you can no longer link against the LearnLib documentation on JDK 8 builds.
- Split packages had to be refactored.
  - The de.learnlib.oracle.parallelism interfaces in the learnlib-api artifact have been moved to the de.learnlib.oracle package.
  - The ThreadSafe caches have been moved from the learnlib-parallelism artifact to the learnlib-cache artifact.
  - The GrowingAlphabet and Resumable tests from the learnlib-learner-it-support artifact been moved to the learnlib-test-support artifact.
  - The OTUtils class no longer provides the displayHTMLInBrowser methods in order to not depend on java.desktop. If you relied on this functionality, use the writeHTMLToFile methods instead and call Desktop.getDesktop().open(file.toURI()) yourself.
  - The classes in the learnlib-learning-examples artifact have their package renamed to de.learnlib.testsupport.example.
With the removal of the learnlib-annotation-processor artifact (see below), the learnlib-build-config artifact is now part of the de.learnlib group again.
The ReuseOracleBuilder and ReuseTreeBuilder classes are now auto-generated and therefore reside in the respective packages of their previously enclosing classes.

Removed

The de.learnlib.tooling:learnlib-annotation-processor artifact has been dropped. The functionality has been moved to a standalone project.
The de.learnlib:learnlib-rpni-edsm and de.learnlib:learnlib-rpni-mdl artifacts have been dropped. The code has been merged with the de.learnlib:learnlib-rpni artifact.
PropertyOracles can no longer set a property. This value is now immutable and must be provided during instantiation. Previously, the internal state wasn't updated accordingly if a property was overridden.
SymbolQueryOracles (and related code such as the respective caches, counters, etc.) have been removed without replacement. Equivalent functionality on the basis of the new AdaptiveMembershipOracles is available instead.

[0.17.0] - 2023-11-15

Full changelog

Added

Migrated the AAAR algorithm from the old closed-source LearnLib (thanks to Markus Frohme).
Added Moore versions of the OP, TTT, and LStar learners (thanks to Mohamad Bayram).
Added the OML (optimal-MAT-learner) active learning algorithm (thanks to Falk Howar).
Added the OSTIA passive learning algorithm (thanks to Aleksander Mendoza-Drosik).
The RPNI learner now supports MooreMachines (thanks to Markus Frohme).
Added new learning algorithms for procedural systems such as SPAs, SBAs, and SPMMs (thanks to Markus Frohme).

Changed

Refactorings
- Many LearnLib packages have been refactored from plural-based keywords to singular-based keywords. Some examples are
  - renamed all learning algorithm packages from de.learnlib.algorithms.* to de.learnlib.algorithm.*.
  - renamed de.learnlib.counterexamples.* to de.learnlib.counterexample.*.
  - renamed de.learnlib.drivers.* to de.learnlib.driver.*.
  - renamed de.learnlib.util.statistics.* to de.learnlib.util.statistic.*.
  - etc.
  While this may cause some refactoring, it should only affect import statements as the names of most classes remain identical.
- Some actual re-namings concern
  - All code concerning visibly push-down automata now uses the "vpa" acronym (previously "vpda"). This includes package names, class names and (Maven) module names.
  - The "discrimination-tree" learner has been renamed to "observation-pack". This includes classes (DTLearnerDFA -> OPLearnerDFA, etc.), package names, and Maven modules. The same refactoring happened for the VPA-based version of the learner.
  - The learnlib-acex Maven module has been merged with the learnlib-counterexamples module.
  - Classes in the learnlib-api have been moved from de.learnlib.api to de.learnlib.
  - Refactored the package de.learnlib.datastructure.pta.pta.* to de.learnlib.datastructure.pta.*.
  - Refactored the package de.learnlib.driver.util.* to de.learnlib.driver.simulator.*.
  - Moved classes from the package de.learnlib.mapper.api.* to de.learnlib.sul.*.
  - Renamed PassiveLearnerVariantTICase to PassiveLearnerVariantITCase.
AbstractTTTHypothesis has received an additional type parameter for its state type.
AutomatonOracle#accepts no longer has a length parameter. Provide a correctly sized input iterable instead.
Classes revolving around the ContextExecutableInputSUL have been moved from the learnlib-mapper module to the learnlib-drivers-basic module.
The CounterOracle and JointCounterOracle have been merged. Now there only exists a single CounterOracle that counts both the number of queries and the number of symbols therein.
The {DFA,Mealy}CacheOracles and the SULCache are no longer thread-safe because the intended pipeline of a parallel setup (as suggested by the LearnLib factory methods) consists of a single-threaded cache that delegates to parallel (non-cached) oracles. Here, the synchronization logic only adds unnecessary overhead. In case you want a shared, thread-safe cache (which was currently not possible to set up conveniently) the learnlib-parallelism module now contains the ThreadSafe{DFA,Mealy,SUL}Caches factories which allow one to construct parallel oracles (whose parameters and return types are tailored towards using our ParallelOracleBuilders factory) with a shared cache. See the in-tree ParallelismExample2 for reference.
PTAs now read their sample inputs as IntSeqs.
PassiveLearningAlgorithm#computeModel did not specify whether repeated calls to the method should yield identical models. It is now explicitly left open to the respective implementation to support this behavior. BlueFringeRPNI{DFA,Mealy,Moore} explicitly does not support this behavior, as the internal prefix-tree acceptor is now constructed on-the-fly as samples are added via the addSample methods. This allows to drop the previously redundant caching of samples and reduce memory pressure. BlueFringeEDSMDFA and BlueFringeMDLDFA still have to cache the samples internally and therefore still support repeated model construction.
The Resumable semantics have changed: the returned state object no longer implements Serializable. We never fully supported the semantics of the interface and never intended to do so. In fact, the old approach failed miserably if any class was involved where we missed an "implements Serializable" statement. In order to prevent confusion by promising false contracts, implementing this markup interface has been removed. Serialization should now be done in user-land via one of the many external (and more optimizable) serialization frameworks such as FST, XStream, etc. See the in-tree ResumableExample for reference.
The ADT class is no longer initialized with a leafSplitter but the extendLeaf and splitLeaf methods take an additional argument. This allows for a more customizable behavior.
The automaton-specific SimulatorOracles are now generated automatically and therefore reside in the package de.learnlib.oracle.membership rather that being an inner-class of the SimulatorOracle.
SymbolQueryCache now needs to be created via the MealyCaches factory.
SimplePOJOTestDriver no longer uses a mapper to suppress SULExceptions but instead operates directly on the POJO with simplified inputs/outputs and propagates any exceptions thrown. To complement this change, the old SimplePOJODataMapper has been renamed to SimplePOJOExceptionMapper and only deals with mapping exceptions now. The old behavior can be restored by combining the two classes manually via SULMappers#apply.
Switched to AutomataLib 0.11.0.

Removed

Removed (unused) de.learnlib.datastructure.pta.pta.PropertyConflictException, de.learnlib.datastructure.observationtable.InvalidRowException.
Removed the (protected) exposeInternalHypothesis method on AbstractAutomatonLStar. Subclasses should directly implement the getHypothesisModel method.
Removed the EquivalenceQueries factory. All provided equivalence checkers are available via public constructors which allow for more flexible parameterization.
Removed the Filter interface and FilterChain class. Instantiating the filters that are required for constructing a filter chain already requires setting the delegate oracles. Therefore, the only effect of this interface is that the attributes of the filters cannot be final.
Removed LearnLogger. All code was migrated to use the native SLF4j facade and now uses the markers provided in the Category class to accentuate the different log messages and provide client code with a means to handle the different log messages (similar to the previous purpose-specific log methods).

Fixed

Fixed a bug when adding new alphabet symbols to LStar-based learners which use a counterexample handler that requires consistency checks.

0.16.0 - 2020-10-12

Full changelog

Added

A StateLocalInputSULSymbolQueryOracle to wrap a StateLocalInputSULin a SymbolQueryOracle (e.g. to learn partial systems with the ADT learner).
Added an example for parallel setups (ParallelismExample1).

Changed

The SULSymbolQueryOracle now better handles the pre and post cycles of a SUL (e.g. calls to the reset method now allow to close the embedded SUL from outside).
Reworked parallel oracles
- Several SULOracle variants are no longer thread-safe. This reduces overhead for scenarios where no parallelism is required.
- The ParallelOracleBuilders factory now offers builder methods for SULs, ObservableSULs, StateLocalInputSULs, MembershipOracless and OmegaMembershipOracles to allow an easy (and correct) construction of parallel setups given one of the mentioned implementations.
Refactored the following packages/classes:
- de.learnlib.oracle.parallelism.ParallelOracleInterruptedException -> de.learnlib.api.oracle.parallelism.BatchInterruptedException
The initialPrefixes and initialSuffixes methods of AbstractExtensibleAutomatonLStar are now final since these values can be provided via the constructor of the class. This allows one to simplify subclasses.
Updated to AutomataLib 0.10.0

Removed

Removed the learnlib.queries.parallel.threshold property. Learning setups that want to use parallelism now need to explicitly setup parallel oracles.
Removed MQUtil#answerQueries{Auto,Parallel} and MQUtil#answerOmegaQueries{Auto,Parallel}.
LassoOracle#isOmegaCounterExample(boolean) has been removed. This decision can be directly integrated into the #findCounterExample method which has more information available.

Fixed

Fixed a bug where NL* would create non-canonical hypotheses (#70, thanks to Joshua Moerman)

0.15.0 - 2020-02-06

Full changelog

Changed

The {DFA,Mealy,}W{p,}MethodEQOracle(MembershipOracle, int, int) constructor no longer interprets its second int parameter as the batch size, but as an estimate for the expected SUL size. In order to explicitly set the batch size of the oracle, use the {DFA,Mealy,}W{p,}MethodEQOracle(MembershipOracle, int, int, int) constructor. Now, the two parameters lookahead and expectedSize will determine the length of the middle part via Math.max(lookahead, expectedSize - hypothesis.size()). This allows to dynamically adjust the length of the middle part throughout the learning process. See LearnLib/automatalib#32.
Several DFA/Mealy specific (oracle) subclasses are now automatically generated. As a result they are no longer an inner class but an independent top-level class. This requires to update the import statements.
JSR305 annotations have been replaced with checker-framework annotations.
- LearnLib (incl. AutomataLib) now follows checker-framework's convention that (non-annotated) types are usually considered non-null unless explicitly annotated with @Nullable.
- LearnLib (incl. AutomataLib) no longer has a (runtime-) dependency on JSR305 (and other javax.*) annotations or includes them in the distribution artifact. This now makes LearnLib (incl. AutomataLib) compliant with Oracle's binary code license and allows LearnLib (incl. AutomataLib) artifacts as-is to be bundled in binary distributions with Oracle's JDKs/JREs.
A lot of code for inferring partial Mealy machines (esp. PartialLStarMealy and PartialObservationTable) has been removed/refactored. The concept of state local inputs is now implemented as a SUL filter and introduces a special StateLocalInputSULOracle which early-answers queries that would traverse unavailable inputs with a previously specified symbol. This way, queries that would traverse undefined input symbols still won't be executed on the SUL but the SUL appears as a 'total' Mealy system to the learner, allowing one to use every currently existing Mealy learner as-is. See the in-tree examples for more information.
SULCache no longer implements MembershipOracle.
Updated to AutomataLib 0.9.0

Removed

As a remainder of its initial implementation, the TTTEventListener (and the corresponding event feature in the TTT algorithm) has been removed due to the lack of usage.

0.14.0 - 2019-02-18

Full changelog

Added

Added support for black-box-checking (thanks to Jeroen Meijer).
Added support for learning partial Mealy Machines with state-local inputs via L* (thanks to Maren Geske).
Added support for resumable caches.
DynamicParallelOracles can now be constructed from a collection of independent oracles.
Support for Java 11. Note: LearnLib/AutomataLib still targets Java 8, and thus needs classes provided by this environment (specifically: annotations from javax.annotation). If you plan to use LearnLib/AutomataLib in a Java 11+ environment, make sure to provide these classes. They are not shipped with LearnLib/AutomataLib.

Changed

Refactored the following packages/classes:
- de.learnlib.api.algorithm.feature.SupportsGrowingAlphabet -> net.automatalib.SupportsGrowingAlphabet
- de.learnlib.api.algorithm.feature.ResumableLearner -> de.learnlib.api.Resumable
Some runtime properties for dynamically configuring LearnLib have been renamed. There now exists the LearnLibProperty enum as a single reference point for all available properties.
The node iterators for discrimination trees are now hidden behind the DiscriminationTreeIterators factory.
Parallel Oracles:
- The withDefault* methods have been removed from the {Dynamic,Static}PrallelOracleBuilders. If needed, use the regular with* methods and supply the public default values from {Dynamic,Static}PrallelOracle.
- The new*ParallelOracle methods from the ParallelOracleBuilders factory no longer interpret a single membership oracle parameter as a supplier to a shared oracle, but rather as a single oracle (and thus return a builder for a parallel oracle with fixed pool size).
Adding new symbols to learning algorithms (via the SupportsGrowingAlphabet interface) now requires the learner to be initialized with a GrowingAlphabet instance. This is to make sure that the user has full control over which alphabet instance should be used instead of LearnLib making decisions on behalf of the user.
Discrimination-Tree based Learners (DT, KV, TTT) now batch queries whenever possible, thus allowing to fully utilize parallel oracles.
Also, see the changes in AutomataLib

Fixed

Several bugs detected by our ongoing efforts to write tests.

0.13.1 - 2018-05-11

Full changelog

Fixed

Fixed an out-of-bounds error in a cache implementation
Fixed visibility issues with exportable classes used for the ResumableLearner interface
Fixed an issue when adding a new symbol to a learner and the initial alphabet was already an instance of GrowableAlphabet
General consolidations (typos, wrong documentation, etc.)

0.13.0 - 2018-02-08

Full changelog

Added

Added randomized version of W(p)-Method based equivalence oracles (see #40)
Added the ADT (adaptive distinguishing tree) active learning algorithm
Added two active learning algorithms for visibly pushdown languages.
Added the RPNI (regular positive-negative inference) passive learning algorithm, including EDSM (evidence-driven state merging) and MDL (minimum description length) variants.
Many active learning algorithms now support adding additional alphabet symbols after initial instantiation/starting of the learning process.
Added support for suspending the learning process to a savable / serializable state. The learning process may be resumed from this state at a later point in time.
Added the AbstractTestWordEQOracle class, which allows one to implement custom equivalence oracles solely based on lazy (stream-based) test-word generation. Existing equivalence oracles (as far as possible) have been reworked to extend this class and thus profit from its built-in laziness and batch (parallelization) support.

Changed

Refactored the Maven artifact and Java package structure. Have a look at the List of LearnLib Artifacts for an updated overview of available artifacts. In general, no functionality should have been removed (except of code marked with @Deprecated). The easiest way to migrate your code to the new version is probably by using the Auto-Import feature of your IDE of choice.

The non-trivial refactorings include:
- API methods no longer use wildcards in generic return parameters. This allows your code to not having to deal with them.
- Changes to AutomataLib.
Replaced System.out and JUL logging, with calls to a SLF4j facade.
Code improvements due to employment of several static code-analysis plugins (findbugs, checkstyle, PMD, etc.) as well as setting up continuous integration at Travis CI.