Evolvable mathematical models: A new artificial Intelligence paradigm

Created by W.Langdon from gp-bibliography.bib Revision:1.4420

  author =       "Paul Grouchy",
  title =        "Evolvable mathematical models: A new artificial
                 Intelligence paradigm",
  school =       "Aerospace Science and Engineering, University of
  year =         "2014",
  address =      "Canada",
  month =        nov,
  keywords =     "genetic algorithms, genetic programming, Artificial
                 Intelligence, Artificial Life, Evolutionary
                 Computation, Evolutionary Robotics",
  URL =          "http://hdl.handle.net/1807/68193",
  URL =          "https://tspace.library.utoronto.ca/handle/1807/68193",
  URL =          "https://tspace.library.utoronto.ca/bitstream/1807/68193/1/Grouchy_Paul_201411_PhD_thesis.pdf",
  size =         "150 pages",
  abstract =     "We develop a novel Artificial Intelligence paradigm to
                 generate autonomously artificial agents as mathematical
                 models of behaviour. Agent/environment inputs are
                 mapped to agent outputs via equation trees which are
                 evolved in a manner similar to Symbolic Regression in
                 Genetic Programming. Equations are comprised of only
                 the four basic mathematical operators, addition,
                 subtraction, multiplication and division, as well as
                 input and output variables and constants. From these
                 operations, equations can be constructed that
                 approximate any analytic function. These Evolvable
                 Mathematical Models (EMMs) are tested and compared to
                 their Artificial Neural Network (ANN) counterparts on
                 two benchmarking tasks: the double-pole balancing
                 without velocity information benchmark and the
                 challenging discrete Double-T Maze experiments with
                 homing. The results from these experiments show that
                 EMMs are capable of solving tasks typically solved by
                 ANNs, and that they have the ability to produce agents
                 that demonstrate learning behaviours. To further
                 explore the capabilities of EMMs, as well as to
                 investigate the evolutionary origins of communication,
                 we develop NoiseWorld, an Artificial Life simulation in
                 which inter-agent communication emerges and evolves
                 from initially non-communicating EMM-based agents.
                 Agents develop the capability to transmit their x and y
                 position information over a one-dimensional channel via
                 a complex, dialogue-based communication scheme. These
                 evolved communication schemes are analysed and their
                 evolutionary trajectories examined, yielding
                 significant insight into the emergence and subsequent
                 evolution of cooperative communication. Evolved agents
                 from NoiseWorld are successfully transferred onto
                 physical robots, demonstrating the transferability of
                 EMM-based AIs from simulation into physical reality.",
  notes =        "Supervisor: Gabriele, M.T. D'Eleuterio",

Genetic Programming entries for Paul Grouchy