Cloud Driven Design of a Distributed Genetic Programming Platform

We describe how we design FlexGP, a distributed genetic programming (GP) system to efficiently run on the cloud. The system has a decentralized, fault-tolerant, cascading startup where nodes start to compute while more nodes are launched. It has a peer-to-peer neighbor discovery protocol which constructs a robust communication network across the nodes. Concurrent with neighbor discovery, each node launches a GP run differing in parameterization and training data from its neighbors. This factoring of parameters across learners produces many diverse models for use in ensemble learning.