S
Stochastic Heuristic Optimization (code for IA-308 lecture on metaheuristics). – README.md

README.md

SHO — Stochastic Heuristics Optimization

SHO is a didactic Python framework for implementing metaheuristics (or evolutionary computation, or search heuristics).

Its main objective is to free students from implementing boring stuff and allow them to concentrate on single operator implementation.

The framework implements a simple sensor placement problem and handle metaheuristics manipulating solutions represented as numerical vectors or bitstrings.

Author: Johann Dreo johann.dreo@thalesgroup.com ⓒ Thales group

Executable

The main interface is implemented in snp.py. New algorithms should be integrated within this file and the interface should not be modified. One may add arguments, but not remove or change the contracts of the existing ones.

The file snp_landscape.py is an example that plots the objective function and a greedy search trajectory for a simple problem with only two dimensions.

Architecture

The design pattern of the framework is a functional approach to composition. The goal is to be able to assemble a metaheuristic, by plugging atomic functions in an algorithm template.

Operators

The base of the pattern is a function that contains the main loop of the algorithm, and call other functions called "operators". Example of those algorithms are in the algo module.

For instance, the random algorithm depends on an objective function func, an initialization operator init and a stopping criterion operator again.

Encoding

Some operator do not depend on the way solutions are encoded (like the stopping criterions) and some operators do depend on the encoding. The former are defined in their own modules while the later are defined in the module corresponding to their encoding (either num or bit).

Interface capture

As they are assembled in an algorithm that do not know their internal in advance, an operators needs to honor an interface. For instance, the init operator's interface takes no input parameter and returns a solution to the problem.

However, some operator may need additional parameters to be passed. To solve this problem, the framework use an interface capture pattern.

There is two ways to capture the interface: either with a functional approach, either with an object-oriented approach. You can use the one you prefer, the advice is to use the functional approach when you can implement your operator as a stateless function, and the object-oriented approach when you need àour operator to manage a state.

Function capture

The functional capture helpers are implemented in the make module. Basically, a function in this module capture the operator function's full interface and returns a function having the expected interface of the operator.

The implicit rule is to use positional arguments for mandatory parameters on which the operator is defined, and keyword arguments for parameters which are specific to the operator.

Object-oriented capture

The object-oriented approach does not need helpers, you just need to define a "functor" class, that is, a class which implements the __call__ interface. This special function member allows to call an instance of a class as if it was a function.

The __call__ method should honor the targeted operator interface. To pass fixed parameters, use the __init__ constructor.

There is an example of an operator implemented this way as the steady class in the sho/iters.py file.

Exercises

Setup

Two example algorithms are provided: a random search and a greedy search. Several useful stopping criterions are provided. The corresponding encoding-dependent operators are also provided, for both numeric and bitstring encodings. The snp.py file shows how to assemble either a numeric greedy solver or a bitstring greedy solver.

To setup your own solver, add your algorithm(s) into the algo.py module, then assemble its instance under its name into snp.py. For instance, if you created the annealing algorithm, you will be able to immediatly assemble num_annealing and bit_annealing.

One should be able to call your solvers with python3 snp.py --solver num_annealing, for instance.

List of exercises

Most exercises consists in adding a single function in an existing module (or your own module) and use assemble it in the main executable.

  1. Implement a simulated annealing.
  2. Implement an evolutionary algorithm.
  3. Implement an expected run time empirical cumulative density function.
  4. Implement a simple design of experiment to determine the best solver.
  5. Provide a solver for a competition.