AALpy is a light-weight automata learning library written in Python. You can start learning automata in just a few lines of code.
Whether you work with regular languages or you would like to learn models of (black-box) reactive systems, AALpy supports a wide range of modeling formalisms, including deterministic, non-deterministic, and stochastic automata, as well as deterministic context-free grammars/pushdown automata.
| Automata Type | Supported Formalisms | Algorithms | Features |
|---|---|---|---|
| Deterministic | DFAs Mealy Machines Moore Machines |
L* KV RPNI |
Seamless Caching Counterexample Processing 13 Equivalence Oracles |
| Non-Deterministic | ONFSM Abstracted ONFSM |
L*ONFSM | Size Reduction Trough Abstraction |
| Stochastic | Markov Decision Processes Stochastic Mealy Machines Markov Chains |
L*MDP L*SMM ALERGIA |
Counterexample Processing Exportable to PRISM format Bindings to jALERGIA |
| Pushdown | VPDA/SEVPA | KVVPA | Specification of exclusive call-return pairs |
AALpy enables efficient learning by providing a large set of equivalence oracles, implementing various conformance testing strategies. Active learning is mostly based on Angluin's L* algorithm, for which AALpy supports a selection of optimizations, including efficient counterexample processing caching. However, the recent addition of efficiently implemented KV algorithm requires (on average) much less interaction with the system under learning than L*. In addition, KV can be used to learn Visibly Deterministic Pushdown Automata (VPDA).
AALpy also includes passive automata learning algorithms, namely RPNI for deterministic and ALERGIA for stochastic models. Unlike active algorithms which learn by interaction with the system, passive learning algorithms construct a model based on provided data.
Use the package manager pip to install the latest release of AALpy:
pip install aalpyTo install current version of the master branch (it might contain bugfixes and added functionalities between releases):
pip install https://github.com/DES-Lab/AALpy/archive/master.zipThe minimum required version of Python is 3.6.
Ensure that you have Graphviz installed and added to your path if you want to visualize models.
For manual installation, clone the repo and install pydot (the only dependency).
If you are interested in automata learning or would like to understand the automata learning process in more detail, please check out our Wiki. On Wiki, you will find more detailed examples on how to use AALpy.
Examples.py contains many examples and it is a great starting point.
All automata learning procedures follow this high-level approach:
- Define the input alphabet and system under learning (SUL)
- Choose the equivalence oracle
- Run the learning algorithm
For more detailed examples, check out:
- How to learn Regex with AALpy
- How to learn MQTT with AALpy
- Few Simple Examples
- Interactive Examples
- Examples.py
Examples.py contains examples covering almost the whole of AALpy's functionality, and it is a great starting point/reference. Wiki has a step-by-step guide to using AALpy and can help you understand AALpy and automata learning in general.
Code snipped demonstrating some of AALpy's functionalities
The following snippet demonstrates a short example in which an automaton is either loaded or randomly generated and then learned.
from aalpy.utils import load_automaton_from_file, save_automaton_to_file, visualize_automaton, generate_random_dfa
from aalpy.automata import Dfa
from aalpy.SULs import AutomatonSUL
from aalpy.oracles import RandomWalkEqOracle
from aalpy.learning_algs import run_Lstar, run_KV
# load an automaton
# automaton = load_automaton_from_file('path_to_the_file.dot', automaton_type='dfa')
# or construct it from state setup
dfa_state_setup = {
'q0': (True, {'a': 'q1', 'b': 'q2'}),
'q1': (False, {'a': 'q0', 'b': 'q3'}),
'q2': (False, {'a': 'q3', 'b': 'q0'}),
'q3': (False, {'a': 'q2', 'b': 'q1'})
}
small_dfa = Dfa.from_state_setup(dfa_state_setup)
# or randomly generate one
random_dfa = generate_random_dfa(alphabet=[1,2,3,4,5],num_states=20, num_accepting_states=8)
big_random_dfa = generate_random_dfa(alphabet=[1,2,3,4,5],num_states=2000, num_accepting_states=500)
# get input alphabet of the automaton
alphabet = random_dfa.get_input_alphabet()
# loaded or randomly generated automata are considered as BLACK-BOX that is queried
# learning algorithm has no knowledge about its structure
# create a SUL instance for the automaton/system under learning
sul = AutomatonSUL(random_dfa)
# define the equivalence oracle
eq_oracle = RandomWalkEqOracle(alphabet, sul, num_steps=5000, reset_prob=0.09)
# start learning
# run_KV is for the most part reacquires much fewer interactions with the system under learning
learned_dfa = run_KV(alphabet, sul, eq_oracle, automaton_type='dfa')
# or run L*
# learned_dfa_lstar = run_Lstar(alphabet, sul, eq_oracle, automaton_type='dfa')
# save automaton to file and visualize it
# save_automaton_to_file(learned_dfa, path='Learned_Automaton', file_type='dot')
# or
learned_dfa.save()
# visualize automaton
# visualize_automaton(learned_dfa)
learned_dfa.visualize()
# or just print its DOT representation
print(learned_dfa)To make experiments reproducible, define a random seed at the beginning of your program.
from random import seed
seed(2) # all experiments will be reproducibleAALpy has been used to:
If you use AALpy in your research, please cite us with of the following:
If you have research suggestions or you need specific help concerning your research, feel free to start a discussion or contact edi.muskardin@silicon-austria.com. We are happy to help you and consult you in applying automata learning in various domains.
Pull requests are welcome. For significant changes, please open an issue first to discuss what you would like to change. In case of any questions or possible bugs, please open issues.