Genetic Programming and a LLVM JIT for restricted Python AST expressions

A small intro on the rationale

So I’m working on a Symbolic Regression Machine written in C/C++ called Shine, which is intended to be a JIT for Genetic Programming libraries (like Pyevolve for instance). The main rationale behind Shine is that we have today a lot of research on speeding Genetic Programming using GPUs (the GPU fever !) or any other special hardware, etc, however we don’t have many papers talking about optimizing GP using the state of art compilers optimizations like we have on clang, gcc, etc.

The “hot spot” or the component that consumes a lot of CPU resources today on Genetic Programming is the evaluation of each individual in order to calculate the fitness of the program tree. This evaluation is often executed on each set of parameters of the “training” set. Suppose you want to make a symbolic regression of a single expression like the Pythagoras Theorem and you have a linear space of parameters from 1.0 to 1000.0 with a step of 0.1 you have 10.000 evaluations for each individual (program tree) of your population !

What Shine does is described on the image below:

It takes the individual of the Genetic Programming engine and then converts it to LLVM Intermediate Representation (LLVM assembly language), after that it runs the transformation passes of the LLVM (here is where the true power of modern compilers enter on the GP context) and then the LLVM JIT converts the optimized LLVM IR into native code for the specified target (X86, PowerPC, etc).

You can see below the Shine architecture:

This architecture brings a lot of flexibility for Genetic Programming, you can for instance write functions that could be used later on your individuals on any language supported by the LLVM, what matters to Shine is the LLVM IR, you can use any language that LLVM supports and then use the IR generated by LLVM, you can mix code from C, C++, Ada, Fortran, D, etc and use your functions as non-terminal nodes of your Genetic Programming trees.

Shine is still on its earlier development, it looks a simple idea but I still have a lot of problems to solve, things like how to JIT the evaluation process itself instead of doing calls from Python using ctypes bindings of the JITed trees.

Doing Genetic Programming on the Python AST itself

During the development of Shine, an idea happened to me, that I could use a restricted Python Abstract Syntax Tree (AST) as the representation of individuals on a Genetic Programming engine, the main advantage of this is the flexibility and the possibility to reuse a lot of things. Of course that a shared library written in C/C++ would be useful for a lot of Genetic Programming engines that doesn’t uses Python, but since my spare time to work on this is becoming more and more rare I started to rethink the approach and use Python and the LLVM bindings for LLVM (LLVMPY) and I just discovered that is pretty easy to JIT a restricted set of the Python AST to native code using LLVM, and this is what this post is going to show.

JIT’ing a restricted Python AST

The most amazing part of LLVM is obviously the amount of transformation passes, the JIT and of course the ability to use the entire framework through a simple API (ok, not so simple sometimes). To simplify this example, I’m going to use an arbitrary restricted AST set of the Python AST that supports only subtraction (-), addition (+), multiplication (*) and division (/).

To understand the Python AST, you can use the Python parser that converts source into AST:

>>> import ast
>>> astp = ast.parse("2*7")
>>> ast.dump(astp)
'Module(body=[Expr(value=BinOp(left=Num(n=2), op=Mult(), right=Num(n=7)))])'

What the parse created was an Abstract Syntax Tree containing the BinOp (Binary Operation) with the left operator as the number 2, the right operator as the number 7 and the operation itself as Multiplication(Mult), very easy to understand. What we are going to do to create the LLVM IR is to create a visitor that is going to visit each node of the tree. To do that, we can subclass the Python NodeVisitor class from the ast module. What the NodeVisitor does is to visit each node of the tree and then call the method ‘visit_OPERATOR’ if it exists, when the NodeVisitor is going to visit the node for the BinOp for example, it will call the method ‘visit_BinOp’ passing as parameter the BinOp node itself.

The structure of the class for for the JIT visitor will look like the code below:

# Import the ast and the llvm Python bindings
import ast
from llvm import *
from llvm.core import *
from import *
import llvm.passes as lp

class AstJit(ast.NodeVisitor):
    def __init__(self):

What we need to do now is to create an initialization method to keep the last state of the JIT visitor, this is needed because we are going to JIT the content of the Python AST into a function and the last instruction of the function needs to return what was the result of the last instruction visited by the JIT. We also need to receive a LLVM Module object in which our function will be created as well the closure type, for the sake of simplicity I’m not type any object, I’m just assuming that all numbers from the expression are integers, so the closure type will be the LLVM integer type.

def __init__(self, module, parameters):
    self.last_state = None
    self.module = module
    # Parameters that will be created on the IR function
    self.parameters = parameters
    self.closure_type =
    # An attribute to hold a link to the created function
    # so we can use it to JIT later
    self.func_obj = None

def _create_builder(self):
    # How many parameters of integer type
    params = [self.closure_type] * len(self.parameters)

    # The prototype of the function, returning a integer
    # and receiving the integer parameters
    ty_func = Type.function(self.closure_type, params)

    # Add the function to the module with the name 'func_ast_jit'
    self.func_obj = self.module.add_function(ty_func, 'func_ast_jit')

    # Create an argument in the function for each parameter specified
    for index, pname in enumerate(self.parameters):
        self.func_obj.args[index].name = pname

    # Create a basic block and the builder
    bb = self.func_obj.append_basic_block("entry")
    self.builder =

Now what we need to implement on our visitor is the ‘visit_OPERATOR’ methods for the BinOp and for the Numand Name operators. We will also implement the method to create the return instruction that will return the last state.

# A 'Name' is a node produced in the AST when you
# access a variable, like '2+x+y', 'x' and 'y' are
# the two names created on the AST for the expression.
def visit_Name(self, node):
    # This variable is what function argument ?
    index = self.parameters.index(
    self.last_state = self.func_obj.args[index]
    return self.last_state

# Here we create a LLVM IR integer constant using the
# Num node, on the expression '2+3' you'll have two
# Num nodes, the Num(n=2) and the Num(n=3).
def visit_Num(self, node):
    self.last_state =, node.n)
    return self.last_state

# The visitor for the binary operation
def visit_BinOp(self, node):
    # Get the operation, left and right arguments
    lhs = self.visit(node.left)
    rhs = self.visit(node.right)
    op = node.op

    # Convert each operation (Sub, Add, Mult, Div) to their
    # LLVM IR integer instruction equivalent
    if isinstance(op, ast.Sub):
        op = self.builder.sub(lhs, rhs, 'sub_t')
    elif isinstance(op, ast.Add):
        op = self.builder.add(lhs, rhs, 'add_t')
    elif isinstance(op, ast.Mult):
        op = self.builder.mul(lhs, rhs, 'mul_t')
    elif isinstance(op, ast.Div):
        op = self.builder.sdiv(lhs, rhs, 'sdiv_t')

    self.last_state = op
    return self.last_state

# Build the return (ret) statement with the last state
def build_return(self):

And that is it, our visitor is ready to convert a Python AST to a LLVM IR assembly language, to run it we’ll first create a LLVM module and an expression:

module ='ast_jit_module')
# Note that I'm using two variables 'a' and 'b'
expr = "(2+3*b+33*(10/2)+1+3/3+a)/2"
node = ast.parse(expr)
print ast.dump(node)

Will output:

left=BinOp(left=BinOp(left=Num(n=2), op=Add(), right=BinOp(
left=Num(n=3), op=Mult(), right=Name(id='b', ctx=Load()))), op=Add(),
right=BinOp(left=Num(n=33), op=Mult(), right=Num(n=2))), op=Add(),
right=Num(n=1)), op=Add(), right=Num(n=3)), op=Add(),
right=Name(id='a', ctx=Load())), op=Div(), right=Num(n=2)))])


Now we can finally run our visitor on that generated AST the check the LLVM IR output:

visitor = AstJit(module, ['a', 'b'])
print module

Will output the LLVM IR:

; ModuleID = 'ast_jit_module'

define i32 @func_ast_jit(i32 %a, i32 %b) {
  %mul_t = mul i32 3, %b
  %add_t = add i32 2, %mul_t
  %add_t1 = add i32 %add_t, 165
  %add_t2 = add i32 %add_t1, 1
  %add_t3 = add i32 %add_t2, 1
  %add_t4 = add i32 %add_t3, %a
  %sdiv_t = sdiv i32 %add_t4, 2
  ret i32 %sdiv_t

Now is when the real fun begins, we want to run LLVM optimization passes to optimize our code with an equivalent GCC -O2 optimization level, to do that we create a PassManagerBuilder and a PassManager, the PassManagerBuilder is the component that adds the passes to the PassManager, you can also manually add arbitrary transformations like dead code elimination, function inlining, etc:

pmb =
# Optimization level
pmb.opt_level = 2

pm =

# Run the passes into the module
print module

Will output:

; ModuleID = 'ast_jit_module'

define i32 @func_ast_jit(i32 %a, i32 %b) nounwind readnone {
  %mul_t = mul i32 %b, 3
  %add_t3 = add i32 %a, 169
  %add_t4 = add i32 %add_t3, %mul_t
  %sdiv_t = sdiv i32 %add_t4, 2
  ret i32 %sdiv_t

And here we have the optimized LLVM IR of the Python AST expression. The next step is to JIT that IR into native code and then execute it with some parameters:

ee =
arg_a =, 100)
arg_b =, 42)

retval = ee.run_function(visitor.func_obj, [arg_a, arg_b])
print "Return: %d" % retval.as_int()

Will output:

Return: 197

And that’s it, you have created a AST->LLVM IR converter, optimized the LLVM IR with the transformation passes and then converted it to native code using the LLVM execution engine. I hope you liked =)

Cite this article as: Christian S. Perone, "Genetic Programming and a LLVM JIT for restricted Python AST expressions," in Terra Incognita, 15/08/2012,

Invite: PyCon US 2011 – Genetic Programming in Python

If you are going to PyCon US 2011, I would like to invite you to the talk “Genetic Programming in Python“, the talk will be given by Eric Floehr on March 12th 1:20 p.m. – 2:05 p.m.

Here is the abstract:

Did you know you can create and evolve programs that find solutions to problems? This talk walks through how to use Genetic Algorithms and Genetic Programming as tools to discover solutions to hard problems, when to use GA/GP, setting up the GA/GP environment, and interpreting the results. Using pyevolve, we’ll walk through a real-world implementation creating a GP that predicts the weather.


Genetic Algorithms (GA) and Genetic Programming (GP) are methods used to search for and optimize solutions in large solution spaces. GA/GP use concepts borrowed from natural evolution, such as mutation, cross-over, selection, population, and fitness to generate solutions to problems. If done well, these solutions will become better as the GA/GP runs.

GA/GP has been used in problem domains as diverse as scheduling, database index optimization, circuit board layout, mirror and lens design, game strategies, and robotic walking and swimming. They can also be a lot of fun, and have been used to evolve aesthetically pleasing artwork, melodies, and approximating pictures or paintings using polygons.

GA/GP is fun to play with because often-times an unexpected solution will be created that will give new insight or knowledge. It might also present a novel solution to a problem, one that a human may never generate. Solutions may also be inscrutable, and determining why a solution works is interesting in itself.

PyOhio 2010: Genetic Programming in Python

Eric Floehr (from Intellovations)  kindly sent me the presentation he presented at PyOhio 2010. I think Eric has captured some nice features of Pyevolve which few people use, like DB Adapters, dot plotting, Interactive Mode, Real Time statistics, etc. He also presents an interesting use case where he uses Genetic Programming in order to forecast weather based on some historical data:

Thank you Eric !

Successful pyevolve multiprocessing speedup for Genetic Programming

As we know, Genetic Programming usually requires intensive processing power for the fitness functions and tree manipulations (in crossover operations), and this fact can be a huge problem when using a pure Python approach like Pyevolve. So, to overcome this situation, I’ve used the Python multiprocessing features to implement a parallel fitness evaluation approach in Pyevolve and I was surprised by the super linear speedup I got for a cpu bound fitness function used to do the symbolic regression of the Pythagoras theorem: c = \sqrt{a^2 + b^2}. I’ve used the same seed for the GP, so it has consumed nearly the same cpu resources for both test categories. Here are the results I obtained:


The first fitness landscape I’ve used had 2.500 points and the later had a fitness landscape of 6.400 points, here is the source code I’ve used (you just need to turn on the multiprocessing option using the setMultiProcessing method, so Pyevolve will use multiprocessing when you have more than one single core, you can enable the logging feature to check what’s going on behind the scenes):

from pyevolve import *
import math

rmse_accum = Util.ErrorAccumulator()

def gp_add(a, b): return a+b
def gp_sub(a, b): return a-b
def gp_mul(a, b): return a*b
def gp_sqrt(a):   return math.sqrt(abs(a))

def eval_func(chromosome):
   global rmse_accum
   code_comp = chromosome.getCompiledCode()

   for a in xrange(0, 80):
      for b in xrange(0, 80):
         evaluated     = eval(code_comp)
         target        = math.sqrt((a*a)+(b*b))
         rmse_accum   += (target, evaluated)
   return rmse_accum.getRMSE()

def main_run():
   genome = GTree.GTreeGP()
   genome.setParams(max_depth=4, method="ramped")
   genome.evaluator += eval_func

   ga = GSimpleGA.GSimpleGA(genome, seed=666)
   ga.setParams(gp_terminals       = ['a', 'b'],
                gp_function_prefix = "gp")


   best = ga.bestIndividual()

if __name__ == "__main__":

As you can see, the population size was 800 individuals with a 8% mutation rate and a 100% crossover rate for a simple 20 generations evolution. Of course you don’t need so many points in the fitness landscape, I’ve used 2.500+ points to create a cpu intensive fitness function, otherwise, the speedup can be less than 1.0 due the communication overhead between the processes. For the first case (2.500 points fitness landscape) I’ve got a 3.33x speedup and for the last case (6.400 points fitness landscape) I’ve got a 3.28x speedup. The tests were executed in a 2 cores pc (Intel Core 2 Duo).

Approximating Pi number using Genetic Programming


As many (or very few in the real life haha) people know, today is the Pi Approximation Day ! So it’s time to make a contribution to celebrate this funny day =)

My contribution is to use Python and Pyevolve to approximate Pi number using Genetic Programming approach. I’ve created the functions gp_add(+), gp_sub(-), gp_div(/), gp_mul(*) and gp_sqrt (square root) to use as non-terminals of the GP. The fitness function is very simple too, it simple returns the absolute difference between the Python math.pi and the evaluated individual. I’ve used also a population size of 1k individuals with max tree depth of 8 and the random ephemeral constants as random integers. The best approximation I’ve got while running the GP for about 8 minutes (40 generations) was 3.1416185511, best for 3 digits, you can improve it and let it run for more time to get better approximations.

Here is the formulae I’ve got with the GP (click to enlarge):


And here is the output of the script:

Best (0): 3.1577998365
        Error: 0.0162071829
Best (10): 3.1417973679
        Error: 0.0002047143
Best (20): 3.1417973679
        Error: 0.0002047143
Best (30): 3.1417973679
        Error: 0.0002047143
Best (40): 3.1416185511
        Error: 0.0000258975

- GenomeBase
        Score:                   0.000026
        Fitness:                 15751.020831

        Params:          {'max_depth': 8, 'method': 'ramped'}

        Slot [Evaluator] (Count: 1)
        Slot [Initializator] (Count: 1)
                Name: GTreeGPInitializator - Weight: 0.50
                Doc: This initializator accepts the follow parameters:

      The max depth of the tree

      The method, accepts "grow" or "full"

   .. versionadded:: 0.6
      The *GTreeGPInitializator* function.

        Slot [Mutator] (Count: 1)
                Name: GTreeGPMutatorSubtree - Weight: 0.50
                Doc:  The mutator of GTreeGP, Subtree Mutator

   .. versionadded:: 0.6
      The *GTreeGPMutatorSubtree* function

        Slot [Crossover] (Count: 1)
                Name: GTreeGPCrossoverSinglePoint - Weight: 0.50

- GTree
        Height:                 8
        Nodes:                  21

GTreeNodeBase [Childs=1] - [gp_sqrt]
  GTreeNodeBase [Childs=2] - [gp_div]
    GTreeNodeBase [Childs=2] - [gp_add]
      GTreeNodeBase [Childs=0] - [26]
      GTreeNodeBase [Childs=2] - [gp_div]
        GTreeNodeBase [Childs=2] - [gp_mul]
          GTreeNodeBase [Childs=2] - [gp_add]
            GTreeNodeBase [Childs=2] - [gp_sub]
              GTreeNodeBase [Childs=0] - [34]
              GTreeNodeBase [Childs=2] - [gp_sub]
                GTreeNodeBase [Childs=0] - [44]
                GTreeNodeBase [Childs=0] - [1]
            GTreeNodeBase [Childs=2] - [gp_mul]
              GTreeNodeBase [Childs=0] - [49]
              GTreeNodeBase [Childs=0] - [43]
          GTreeNodeBase [Childs=1] - [gp_sqrt]
            GTreeNodeBase [Childs=0] - [18]
        GTreeNodeBase [Childs=0] - [16]
    GTreeNodeBase [Childs=2] - [gp_add]
      GTreeNodeBase [Childs=0] - [24]
      GTreeNodeBase [Childs=0] - [35]

- GTreeGP
        Expression: gp_sqrt(gp_div(gp_add(26,
gp_sub(44, 1)), gp_mul(49, 43)), gp_sqrt(18)),
16)), gp_add(24, 35)))

And finally, here is the source code:

from __future__ import division
from pyevolve import *
import math

def gp_add(a, b): return a+b
def gp_sub(a, b): return a-b
def gp_div(a, b): return 1 if b==0 else a/b
def gp_mul(a, b): return a*b
def gp_sqrt(a):   return math.sqrt(abs(a))

def eval_func(chromosome):
   code_comp = chromosome.getCompiledCode()
   ret = eval(code_comp)
   return abs(math.pi - ret)

def step_callback(engine):
   gen = engine.getCurrentGeneration()
   if gen % 10 == 0:
      best = engine.bestIndividual()
      best_pi = eval(best.getCompiledCode())
      print "Best (%d): %.10f" % (gen, best_pi)
      print "\tError: %.10f" % (abs(math.pi - best_pi))

   return False

def main_run():
   genome = GTree.GTreeGP()

   genome.setParams(max_depth=8, method="ramped")
   genome.evaluator += eval_func

   ga = GSimpleGA.GSimpleGA(genome)
   ga.setParams(gp_terminals       = ['ephemeral:random.randint(1, 50)'],
                gp_function_prefix = "gp")


   best = ga.bestIndividual()

   print best

if __name__ == "__main__":

If you are interested why today is the Pi Approximation day, see some resources:

Little Cartoon

Some Background History

Some Pi Approximations

Genetic Programming and Flex layouts

To show how Genetic Programming of Pyevolve can be flexible, I’ve done a simple example using Adobe Flex and Pyevolve, the example is just to show how to evolve some kind of Flex layouts, I’ve not implemented the fitness function, this example will just create a random Flex layout using MXML. So, here is the Pyevolve code of the example:

import random
from pyevolve import *

def gp_hbox(x, y):
   return "%s %s" % (x,y)

def gp_vbox(x, y):
   return "%s %s" % (x,y)

def gp_panel(x, y):
   return "%s %s" % (x,y)

def eval_func(chromosome):
   code_comp = chromosome.getCompiledCode()

   for a in xrange(0, 5):
      for b in xrange(0, 5):
         evaluated     = eval(code_comp)
   return random.randint(1,100)

def main_run():
   genome = GTree.GTreeGP()
   genome.setParams(max_depth=5, method="ramped")
   genome.evaluator += eval_func

   ga = GSimpleGA.GSimpleGA(genome)

   button     = repr("<mx:Button label='Button'/>")
   label      = repr("<mx:Label text='Label'/>")
   text_input = repr("<mx:TextInput width='50'/>")

   ga.setParams(gp_terminals       = [button, label, text_input],
                gp_function_prefix = "gp")
   print ga.bestIndividual()

if __name__ == "__main__":

As you can see, I’ve created the layout tags like HBox, VBox and Panel as functions of GP and the Button, Labe, TextInput as terminals of the GP, the result is very funny, it’s just a random layout, but you can use your imagination to create some nice and interesting fitness functions.

Here is the SWF generated from a random individual of the population:

I hope you enjoyed =)