Python abstraction around Binary Decision Diagrams to implement Multi-valued Decision Diagrams.
Table of Contents
Multi-valued Decision Diagrams (MDD) are a way to represent discrete function:
f : A₁ × A₂ × … × Aₙ → B.
Conceptually, a MDD for f can be thought of as a compressed decision tree (in the form of a directed acyclc graph).
For example, if we have a function over two variables,
x ∈ {1,2,3,4,5}, y ∈ {'a','b'}
with possible outputs f(x, y) ∈ {-1, 0, 1}, then the following diagram represents the function:
f(x, y) = 1 if (x ≡ 1 and y ≡ 'a') else 0
This library provides abstractions to easily create and manipulate MDDs.
If you just need to use py-mdd, you can just run:
$ pip install mdd
For developers, note that this project uses the poetry python package/dependency management tool. Please familarize yourself with it and then run:
$ poetry install
For the impatient, here is a basic usage example:
import mdd
interface = mdd.Interface(
inputs={
"x": [1, 2, 3],
"y": [6, 'w'],
"z": [7, True, 8],
},
output=[-1, 0, 1],
)
func = interface.constantly(-1)
assert func({'x': 1, 'y': 'w', 'z': 8}) == -1
# Can access underlying BDD from `dd` library.
# Note: This BDD encodes both the function's output
# *and* domain (valid inputs).
assert func.bdd.dag_size == 33If 33 seems very large to you, this is just a constant function after all, note that as the following sections illustrate, its easy to implement alternative encodings which can be much more compact. [0]
The mdd api centers around three objects:
Variable: Representation of a named variable taking on values in from a finite set described by an aiger circuit.Interface: Description of inputs and outputs of a Multi-valued Decision Diagram.DecisionDiagram: Representation of a Multi-valued Decision Diagram that conforms to an interface. This object is a wrapper around a Binary Decision Diagram object (from dd).
By default, variables use one-hot encoding, but all input
variables can use arbitrary encodings by defining a bit-vector
expression describing valid inputs and a encode/decoder pair from
ints to the variable's domain.
# One hot encoded by default.
var1 = mdd.to_var(domain=["x", "y", "z"], name="myvar1")
# Hand crafted encoding using `py-aiger`.
import aiger_bv
# Named 2-length bitvector circuit.
bvexpr = aiger_bv.uatom(2, 'myvar3')
domain = ('a', 'b', 'c')
var2 = mdd.Variable(
encode=domain.index, # Any -> int
decode=domain.__getitem__, # int -> Any
valid=bvexpr < 4, # 0b11 is invalid!
)
# Can create new variable using same encoding, but different name.
var3 = var2.with_name("myvar3")
var4 = mdd.to_var(domain=[-1, 0, 1], name='output')A useful feature of variables is that they can generate an aiger_bv
BitVector object to describe circuits in terms of a variable.
a_int = var2.encode('a')
y_int = var1.encode('y')
# BitVector Expression testing if var2 is 'a' and var1 is 'y'.
expr = (var2.expr() == a_int) & (var1.expr() == y_int)Given these variables, we can define an input/output interface.
# All variables must have distinct names.
interface = mdd.Interface(inputs=[var1, var2, var3], output=var4)Further, as the first example showed, if the default encoding is fine, then we can simply pass a dictionary inplace of inputs and/or a iterable in place of the output. In this case, a 1-hot encoding will be created using the order of the variables.
interface = mdd.Interface(
inputs={
"x": [1, 2, 3], # These are
"y": [6, 'w'], # 1-hot
"z": [7, True, 8], # Encoded.
},
output=[-1, 0, 1], # uuid based output name.
)Finally, given an interface we can create a Multi-valued Decision
Diagram. There are five main ways to create a DecisionDiagram:
-
Given an interface, create a constant function:
func = interface.constantly(1)
-
Wrap an
py-aigercompatible object:import aiger_bv as BV x = interface.var('x') # Access 'x' variable. out = interface.output # Access output variable. expr = BV.ite( x.expr() == x.encode(2), # Test. out.expr() == out.encode(0), # True branch. out.expr() == out.encode(-1), # False branch. ) func = interface.lift(expr) assert func({'x': 2, 'y': 6, 'z': True}) == 0 assert func({'x': 1, 'y': 6, 'z': True}) == -1
-
Wrap an existing Binary Decision Diagram:
bdd = mdd.to_bdd(expr) # Convert `aiger` expression to bdd. func = interface.lift(bdd) # bdd type comes from `dd` library. assert func({'x': 2, 'y': 6, 'z': True}) == 0 assert func({'x': 1, 'y': 6, 'z': True}) == -1
-
Partially Evaluate an existing
DecisionDiagram:constantly_0 = func.let({'x': 2}) assert func({'y': 6, 'z': True}) == 0
-
Override an existing
DecisionDiagramgiven a predicate:# Can be a BDD or and py-aiger compatible object. test = x.expr() == x.encode(1) # If x = 1, then return 1, otherwise return using func. func2 = func.override(test=test, value=1) assert func2({'x': 1, 'y': 6, 'z': True}) == 1
The py-mdd library uses a Binary Decision Diagram to represent a
multi-valued function. The encoding slighly differs from the standard
reduction [1] from mdds to bdds by assuming the following:
- If a variable encoding is invalid, then the bdd maps it to
0. - The output is 1-hot encoded, i.e., there is a variable for each outcome.
- If a
bddhas all input variables fixed to a valid assignment, the resultingbdddepends on exactly one output varible, which then determines the output.
Any bdd that conforms to this encoding can be wrapped up by an
approriate Interface.
The underlying BDD can be reordered to respect variable ordering by
providing a complete list of variable names to the order method.
func.order(['x', 'y', 'z', func.output.name])If the networkx python package is installed:
$ pip install networkx
or the nx option is added when installing py-mdd:
$ pip install mdd[nx]
then one can export a DecisionDiagram as a directed graph.
note: for now, this graph is only partially reduced. In the future, the plan is to guarantee that the exported DAG is fulled reduced.
from mdd.nx import to_nx
graph = to_nx(func) # Has BitVector expressions on edges to represent guards.
graph2 = to_nx(func, symbolic_edges=False) # Has explicit sets of values on edges to represent guards.[0]: To get a sense for how much overhead is introduced, consider the corresponding Zero Suppressed Decision Diagram (ZDD) of a 1-hot encoding. A classic result (see Art of Computer Programming vol 4a) is the ZDD size bounds the BDD size via O(#variables*|size of ZDD|).
[1]: Srinivasan, Arvind, et al. "Algorithms for discrete function manipulation." 1990 IEEE international conference on computer-aided design. IEEE Computer Society, 1990.
