Abstract
The theory of evolution by natural selection has long served as a powerful metaphor and
mechanism for computational problem-solving. Inspired by biological processes, evolutionary
algorithms and artificial neural networks have become central to the development
of adaptive systems and intelligent behaviour. This dissertation introduces GEP-NEAT, a
novel hybrid algorithm combining the symbolic expressiveness of Gene Expression Programming
(GEP) with the dynamic topology evolution of NeuroEvolution of Augmenting
Topologies (NEAT).
NEAT is known for its ability to evolve neural network structures incrementally, but it
suffers from computational inefficiencies, particularly due to topological sorting during
evaluation. GEP-NN, a GEP-based approach to evolving neural networks, offers an alternative
by representing networks as expression trees, yet lacks the maturity and robustness of
NEAT. GEP-NEAT introduces a new representation scheme where innovation numbers are
encoded as subtree configurations, enabling speciation amongst candidate solutions, and
the modularity and reuse of functional components. This hierarchical encoding facilitates
the evolution of more expressive and scalable neural architectures.
The proposed algorithm was evaluated on benchmark problems including XOR, multiplexer
tasks, and CartPole. On the XOR problem, GEP-NEAT achieved success rates
of 60% (redundant configuration) and 70% (compact configuration), outperforming traditional
GEP-NN in the compact case (30%) and approaching NEAT’s perfect success
rate, albeit with higher average convergence times (107-234 generations versus NEAT’s
32 generations). For the 2-multiplexer problem, a multigenic configuration achieved a
75% success rate, compared to 40% for an unigenic configuration, while converging in
fewer generations on average (426 versus 526) and reducing runtime by approximately...