Automated experimentation using active learning

In Science, experiments are empirical observations selected for the
arbitration of competing hypotheses. The research aims to answer
the question of whether machines can select near-optimal efficient
trial sequences for experimentation. The work will be conducted in
the area of automatically suggesting trials for gathering empirical
data for machine learning of agent strategies of animal behaviour.
Trials to provide data for learning involve cost, related to time,
material costs and salaries of experimenters. To that extent, the
efficiency of the experimental process relies on the number of experiments
performed and their associated costs. In this PhD the student will
study how the cost of experimentation can be reduced using automated
trial selection to support machine learning of agent strategies. Previous
automated experimentation has not studied the machine conjecture of
agent strategies. Another novelty of this research lies in its use
of state-of-the-art Bayesian analysis within the Meta-Interpretive
Learning framework. Within this work we allocate a Bayesian posterior
distribution over the hypothesis space. At each iteration, the learner
conducts an experiment which provides the label of an instance having
maximum entropy with respect to the space of hypotheses. Such experiments
should produce near maximal discrimination over the remaining competing
hypotheses, and thus achieve near-maximal reduction of the hypothesis
space. The student will develop the theoretical framework, implement
a working experimentation system and evaluate the gain on the cost
of experimentation for the tasks of learning agent strategies for
insect models. Data and experiment for the insect models will be provided
by collaboration with ecological researchers at the University of
Reading. The research is aligned with the EPSRC research area "Artificial
intelligence technologies".