Poodle - AI Planning in Python

Poodle is the Python-to-PDDL compiler and automated programming framework in an early stage of development.

Rationale

PDDL is a widely-used language to describe AI planning domains. The applications include various robotic planning problems, scheduling, logistics and manufacturing optimization, writing intelligent agents in computer games, real-time decision making, and even automated unix administration. AI planning, and specifically model-based planning, can be explained as a problem-solving method where the software developer describes (models) a problem, rather than codes the algorithm to solve the problem - which is radically different from how the conventional software development is practically-always done today. Not having to invent and code the algorithm has obvious benefits: developer productivity goes to extremes, you can write software with humanly-impossible complexity of algorithms, any tasks that require combining actions into meaningful chains can now be automated.

But despite these extreme gains, AI planning-based software is virtually nonexistent. And there are reasons why imperative programming is so popular and logic programming is not. Imperative programming has a much lower barrier of entry. Realistically, the majority of problems are much easier to code in a "usual" imperative way rather than modeling the full domain. The tooling, ecosystem, coding paradigms, and the language itself are much more polished and well-designed. Finally, many software libraries and components were written, and are readily available, in imperative programming languages.

Poodle aims to change that. The goal is to create a "native merge" of Python and model-based planning. This means that the developer will have an option to either write the algorithm or describe the problem and let the AI figure out the algorithm - with the result as usable in both options. The goal is to develop all the necessary tooling to enable full-scale production use of AI planning in real-world computing tasks - building on the top of a strong foundation created by the Python community.

Translating full Python programs into planning domain enables the use of efficient search methods to compose pre-built Python libraries into new algorithms. And a developer always gets an alternative to use the code imperatively - whenever she desires to switch.

Quickstart

$ pip install poodle # needs Python 3.7+

Let's say you have:

from poodle import Object, xschedule

class World(Object): 
    prepared: int
    said: bool 

def hello(world: World):
    assert world.said == False
    world.prepared += 1

def world(world: World):
    assert world.prepared > 0
    world.said = True
    return "Hello, world!"

w = World()
w.prepared = 0
w.said = False

Now you have two options:

  1. (obvious) execute natively, if you know the algorithm
hello(w)
print(world(w)) 
# -> "Hello, World!"
  1. if you don't know the parameters and/or sequence of execution - ask AI to figure out
print(xschedule(methods=[world, hello], space=[w], goal=lambda:w.said==True))
# -> "Hello, World!"

This will run the code on a hosted solver. To run a local solver, please scroll down to Installation section.

Overview

Introduction

Poodle is a Python module that enables construction of complex planning and constraint satisfaction problems using familiar Pythonic paradigms in production environments. It is still in the early stage of development, but is already powering kubectl‑val, our tool to prevent Kubernetes configuration errors.

Poodle introduces a pair of Python functions called xschedule and schedule that implement an automated planning mechanism, and a new base object Object:

xschedule(
    methods=[...],   # methods
    space=[...],     # objects
    goal=lambda: ... # condition for final object state
)

where methods is the list of methods that the planner should use to try to reach the goal state; space contains the list of Object objects that the planner will try to use as parameters for the methods, and goal is a simple end-state condition expressed as Python logical expression, usually a lambda function.

Object is a special object type that knows how to translate itself to PDDL.

To understand how to construct a problem, let's start with a classic "Hello, World" function:

from poodle import Object, xschedule

class World(Object): # a class that defines object that will hold final state
    said: bool       # declaration of a bollean variable (Python 3 type hints)

def hello(world: World): # annotated function that mutates the state of `world`
    assert world.said == False # hint for the planner when this call is valid
    print("Hello, World!")
    world.said = True    # mutate the state of the parameter object

w = World()          # create first object instance
w.said = False       # define the value for `said` attribute

# now execute this in an unfamiliar way ... 
xschedule(methods=[hello], space=[w], goal=lambda:w.said==True)

This program will immediately print "Hello, World!" to the console, which looks obvious at first. What actually happened is that Poodle compiled your Python method into PDDL domain + problem and used AI planner to find that the final state is achievable by simply executing the only method, and all asserts are satisfied with our hero object w.

It is important to note that the more precisely you describe your task, the easier it is for the AI planner to figure out the algorithm. That is why Poodle enforces fully statically typed interface for all objects and methods in search space as a minimum selectivity requirement. This also saves you from a lot of bugs in bigger projects.

Let's now jump to a more sophisticated example:

Monkey and Banana problem

monkey-banana-problem

We need to plan monkey actions to get the hanging banana: move the box, climb the box and grasp the bananas. We also need to take into account some basic laws of nature: like if the banana is on the tree, it's location is where the tree is.

from poodle import Object, schedule
from typing import Set

class Position(Object):
    def __str__(self):
        if not hasattr(self, "locname"): return "unknown"
        return self.locname
class HasHeight(Object):
    height: int
class HasPosition(Object):
    at: Position
class Monkey(HasHeight, HasPosition): pass
class PalmTree(HasHeight, HasPosition): 
    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        self.height = 2
class Box(HasHeight, HasPosition): pass
class Banana(HasHeight, HasPosition): 
    owner: Monkey
    attached: PalmTree 
class World(Object):
    locations: Set[Position]


p1 = Position()
p1.locname = "Position A"
p2 = Position()
p2.locname = "Position B"
p3 = Position()
p3.locname = "Position C"

w = World()
w.locations.add(p1)
w.locations.add(p2)
w.locations.add(p3)

m = Monkey()
m.height = 0 # ground
m.at = p1

box = Box()
box.height = 2
box.at = p2

p = PalmTree()
p.at = p3

b = Banana()
b.attached = p

def go(monkey: Monkey, where: Position):
    assert where in w.locations
    assert monkey.height < 1, "Monkey can only move while on the ground"
    monkey.at = where
    return f"Monkey moved to {where}"

def push(monkey: Monkey, box: Box, where: Position):
    assert monkey.at == box.at
    assert where in w.locations
    assert monkey.height < 1, "Monkey can only move the box while on the ground"
    monkey.at = where
    box.at = where
    return f"Monkey moved box to {where}"

def climb_up(monkey: Monkey, box: Box):
    assert monkey.at == box.at
    monkey.height += box.height
    return "Monkey climbs the box"

def grasp(monkey: Monkey, banana: Banana):
    assert monkey.height == banana.height
    assert monkey.at == banana.at
    banana.owner = monkey
    return "Monkey takes the banana"

def infer_owner_at(palmtree: PalmTree, banana: Banana):
    assert banana.attached == palmtree
    banana.at = palmtree.at
    return "Remembered that if banana is on palm tree, its location is where palm tree is"

def infer_banana_height(palmtree: PalmTree, banana: Banana):
    assert banana.attached == palmtree
    banana.height = palmtree.height
    return "Remembered that if banana is on the tree, its height equals tree's height"

print('\n'.join(x() for x in schedule(
          [go, push, climb_up, grasp, infer_banana_height, infer_owner_at],
          [w,p1,p2,p3,m,box,p,b],
          goal=lambda: b.owner == m)))

this program solves the slightly modified "Monkey and banana" planning problem and produces the result:

$ pip install poodle
$ python ./monkey.py
Monkey moved to Position B
Remembered that if banana is on the tree, its height equals tree's height
Remembered that if banana is on palm tree, its location is where palm tree is
Monkey moved box to Position C
Monkey climbs the box
Monkey takes the banana

Kubernetes Configuration Problem

kubernetes-evict-problem

In this example, we are checking what are the consequences of configuration changes in a Kubernetes cluster. For example, the Kubernetes current state may be that once you load a new DaemonSet, important services will get evicted from the cluster in order to place your new microservice.

For a complete program, feel free to check out kubectl-val source code.

Principles and Architecture

Poodle compiles Python into PDDL and uses fast-downward to run the search. As a typical real-world problem requires huge amounts of RAM, the whole solver bundle is running as an HTTP service in current architecture.

Composability

Support for nested xschedule is on the roadmap for planning code composability, although Python already provides excellent composability mechanisms.

Readability and Debuggability

Bad readability and debuggability have always plagued logic languages, and Poodle is not an exception: it is hard to tell what the result would be just by reading the code, as multiple methods can be executed concurrently and in any order. To address this problem, adding a visual debugger based on VOWL is planned. Although a combination of good code design and classical Python REPL, plus some CLIPS inferencer tricks, allowed us to rapidly develop quite sophisticated AI planning-based software.

Documentation

There is no documentation at this point, but we promise to provide it as poodle evolves. If you would like to experiment with Poodle, the general recommendation is to start from reading the examples, unit tests and the kubectl-val project source.

Installation

pip install poodle

Poodle requires Python 3.7+ and will drop support for Python 3.7 as soon as 3.8 is stable due to heavy use of type hinting features.

Running Local Solver

By default, Poodle will check if local port 16009 is open and use a solver running on localhost. If it can not find a local solver it will use a hosted solver environment from CriticalHop, which has some limitations in its free-to-use version.

To run a local solver, you must first install fast-downward. After you have fast-downward running, run poodleserver (included with poodle) from the fast-downward folder:

cd fast-downward
poodleserver

You can also specify the solver URL by environment variable POODLE_SOLVER_URL, e.g.:

export POODLE_SOLVER_URL=http://localhost:8082

Problem Sharing

If you would like to support development of AI planners, we kindly ask you to opt-in for sharing of anonymized PDDL problem data sets. This will help us to continuously improve result waiting times for everyone in the community. To enable this, please set POODLE_STATS=1 environment variable when launching poodleserver. Your privacy is our highest priority, therefore we only collect generated anonymized PDDL data-sets.

Developing

Poodle development is set up with tox and poetry. To run all tests locally, compile fast-downward in a folder named downward near your poodle clone, then do

POODLE_SOLVER_URL=http://localhost:12345 tox

where 12345 is the port of your choosing. This command will run local solver and all tests sequentially. You may then use poetry to manage your local installation.

The correct folder structure for poodle development is:

poodle-dev/
    downward/ -> Fast-Downward installation
    poodle/   -> Poodle repository clone

GitHub