Genetic programming

Not to be confused with Genetic engineering.

In artificial intelligence, genetic programming (GP) is an evolutionary algorithm-based methodology inspired by biological evolution to find computer programs that perform a user-defined task. It is a specialization of genetic algorithms (GA) where each individual is a computer program. It is a machine learning technique used to optimize a population of computer programs according to a fitness landscape determined by a program's ability to perform a given computational task.

Contents

History

In 1954, GP began with the evolutionary algorithms first used by Nils Aall Barricelli applied to evolutionary simulations. In the 1960s and early 1970s, evolutionary algorithms became widely recognized as optimization methods. Ingo Rechenberg and his group were able to solve complex engineering problems through evolution strategies as documented in his 1971 PhD thesis and the resulting 1973 book. John Holland was highly influential during the 1970s.

In 1964, Lawrence J. Fogel, one of the earliest practitioners of the GP methodology, applied evolutionary algorithms to the problem of discovering finite-state automata. Later GP-related work grew out of the learning classifier system community, which developed sets of sparse rules describing optimal policies for Markov decision processes. The first statement of modern "tree-based" Genetic Programming (that is, procedural languages organized in tree-based structures and operated on by suitably defined GA-operators) was given by Nichael L. Cramer (1985).[1] This work was later greatly expanded by John R. Koza, a main proponent of GP who has pioneered the application of genetic programming in various complex optimization and search problems.[2]

In the 1990s, GP was mainly used to solve relatively simple problems because it is very computationally intensive. Recently GP has produced many novel and outstanding results in areas such as quantum computing, electronic design, game playing, sorting, and searching, due to improvements in GP technology and the exponential growth in CPU power.[3] These results include the replication or development of several post-year-2000 inventions. GP has also been applied to evolvable hardware as well as computer programs.

Developing a theory for GP has been very difficult and so in the 1990s GP was considered a sort of outcast among search techniques. But after a series of breakthroughs in the early 2000s, the theory of GP has had a formidable and rapid development. So much so that it has been possible to build exact probabilistic models of GP (schema theories, Markov chain models and meta-optimization algorithms).

Chromosome representation

GP evolves computer programs, traditionally represented in memory as tree structures.[4] Trees can be easily evaluated in a recursive manner. Every tree node has an operator function and every terminal node has an operand, making mathematical expressions easy to evolve and evaluate. Thus traditionally GP favors the use of programming languages that naturally embody tree structures (for example, Lisp; other functional programming languages are also suitable).

Non-tree representations have been suggested and successfully implemented, such as linear genetic programming which suits the more traditional imperative languages [see, for example, Banzhaf et al. (1998)]. The commercial GP software Discipulus, uses AIM, automatic induction of binary machine code[5] to achieve better performance. µGP[6] uses a representation similar to linear GP to generate programs that fully exploit the syntax of a given assembly language.

Genetic operators

The main operators used in evolutionary algorithms such as GP are crossover and mutation.

Crossover

Crossover is applied on an individual by simply switching one of its nodes with another node from another individual in the population. With a tree-based representation, replacing a node means replacing the whole branch. This adds greater effectiveness to the crossover operator. The expressions resulting from crossover are very much different from their initial parents.

Mutation

Mutation affects an individual in the population. It can replace a whole node in the selected individual, or it can replace just the node's information. To maintain integrity, operations must be fail-safe or the type of information the node holds must be taken into account. For example, mutation must be aware of binary operation nodes, or the operator must be able to handle missing values.

MOSES

Meta-Optimizing Semantic Evolutionary Search (MOSES) is a meta-programming technique for evolving programs by iteratively optimizing genetic populations.[7] It has been shown to strongly outperform genetic and evolutionary program learning systems, and has been successfully applied to many real-world problems, including computational biology, sentiment evaluation, and agent control.[8] When applied to supervised classification problems, MOSES performs as well as, or better than support vector machines (SVM), while offering more insight into the structure of the data, as the resulting program demonstrates dependencies and is understandable in a way that a large vector of numbers is not.[9]

MOSES is able to out-perform standard GP systems for two important reasons. One is that it uses estimation of distribution algorithms (EDA) to determine the Markov blanket (that is, the dependencies in a Bayesian network) between different parts of a program. This quickly rules out pointless mutations that change one part of a program without making corresponding changes in other, related parts of the program. The other is that it performs reduction to reduce programs to normal form at each iteration stage, thus making programs smaller, more compact, faster to execute, and more human readable. Besides avoiding spaghetti code, normalization removes redundancies in programs, thus allowing smaller populations of less complex programs, speeding convergence.

Other approaches

The basic ideas of genetic programming have been modified and extended in a variety of ways:

Meta-Genetic Programming

Meta-Genetic Programming is the proposed meta learning technique of evolving a genetic programming system using genetic programming itself. It suggests that chromosomes, crossover, and mutation were themselves evolved, therefore like their real life counterparts should be allowed to change on their own rather than being determined by a human programmer. Meta-GP was formally proposed by Jürgen Schmidhuber in 1987,[10] but some earlier efforts may be considered instances of the same technique, including Doug Lenat's Eurisko. It is a recursive but terminating algorithm, allowing it to avoid infinite recursion.

Critics of this idea often say this approach is overly broad in scope. However, it might be possible to constrain the fitness criterion onto a general class of results, and so obtain an evolved GP that would more efficiently produce results for sub-classes. This might take the form of a Meta evolved GP for producing human walking algorithms which is then used to evolve human running, jumping, etc. The fitness criterion applied to the Meta GP would simply be one of efficiency.

For general problem classes there may be no way to show that Meta GP will reliably produce results more efficiently than a created algorithm other than exhaustion. The same holds for standard GP and other search algorithms.

Implementations

Possibly most used:

Others:

See also

References and notes

  1. ^ Nichael Cramer's HomePage
  2. ^ genetic-programming.com-Home-Page
  3. ^ humancompetitive
  4. ^ Cramer, 1985
  5. ^ (Peter Nordin, 1997, Banzhaf et al., 1998, Section 11.6.2-11.6.3)
  6. ^ MicroGP page from CAD Group, Politecnico di Torino
  7. ^ OpenCog MOSES
  8. ^ Moshe Looks (2006), Competent Program Learning, PhD Thesis, http://metacog.org/doc.html 
  9. ^ ibid.
  10. ^ 1987 THESIS ON LEARNING HOW TO LEARN, METALEARNING, META GENETIC PROGRAMMING, CREDIT-CONSERVING MACHINE LEARNING ECONOMY

Bibliography

External links