Astro Teller

Turing completeness in the language of genetic programming with indexed memory

Genetic programming is a method for evolving functions that find approximate or exact solutions to problems. There are many problems that traditional genetic programming (GP) cannot solve, due to the theoretical limitations of its paradigm. A Turing machine (TM) is a theoretical abstraction that expresses the extent of the computational power of algorithms. Any system that is Turing complete is sufficiently powerful to recognize all possible algorithms. GP is not Turing complete. This paper proves that when GP is combined with the technique of indexed memory, the resulting system is Turing complete. This means that, in theory, GP with indexed memory can be used to evolve any algorithm.<<ETX>>

A Contolled Experiment: Evolution for Learning Difficult Image Classification

The signal-to-symbol problem is the task of converting raw sensor data into a set of symbols that Artificial Intelligence systems can reason about. We have developed a method for directly learning and combining algorithms that map signals into symbols. This new method is based on evolutionary computation and imposes little burden on or bias from the humans involved. Previous papers of ours have focused on PADO, our learning architecture. We showed how it applies to the general signal-to-symbol task and in particular the impressive results it brings to natural image object recognition. The most exciting challenge this work has received is the idea that PADOs success in natural image object recognition may be due to the underlying simplicity of the problems we posed it. This challenge implicitly assumes that our approach suffers from many of the same afflictions that traditional computer vision approaches suffer in natural image object recognition. This paper responds to this challenge by designing and executing a controlled experiment specifically designed to solidify PADOs claim to success.

Internal reinforcement in a connectionist genetic programming approach

Abstract Genetic programming (GP) can learn complex concepts by searching for the target concept through evolution of a population of candidate hypothesis programs. However, unlike some learning techniques, such as Artificial Neural Networks (ANNs), GP does not have a principled procedure for changing parts of a learned structure based on that structures performance on the training data. GP is missing a clear, locally optimal update procedure, the equivalent of gradient-descent backpropagation for ANNs. This article introduces a new algorithm, “internal reinforcement”, for defining and using performance feedback on program evolution. This internal reinforcement principled mechanism is developed within a new connectionist representation for evolving parameterized programs, namely “neural programming”. We present the algorithms for the generation of credit and blame assignment in the process of learning programs using neural programming and internal reinforcement. The article includes a comprehensive overview of genetic programming and empirical experiments that demonstrate the increased learning rate obtained by using our principled program evolution approach.