Exploration and Exploitation with Insufficient Resources
Proceedings of the Workshop on On-line Trading of Exploration and Exploitation 2, PMLR 26:37-61, 2012.
In physical experimentation, the resources available to discover new knowledge are typically extremely small in comparison to the size and dimensionality of the parameter spaces that can be searched. Additionally, due to the nature of physical experimentation, experimental errors will occur, particularly in biochemical experimentation where the reactants may undetectably denature, or reactant contamination could occur or equipment failure. These errors mean that not all experimental measurements and observations will be accurate or representative of the system being investigated. As the validity of observations is not guaranteed, resources must be split between exploration to discover new knowledge and exploitation to test the validity of the new knowledge. Currently we are investigating the automation of discovery in physical experimentation, with the aim of producing a fully autonomous closed-loop robotic machine capable of autonomous experimentation. This machine will build and evaluate hypotheses, determine experiments to perform and then perform them on an automated lab-on-chip experimentation platform for biochemical response characterisation. In the present work we examine how the trade-off between exploration and exploitation can occur in a situation where the number of experiments that can be performed is extremely small and where the observations returned are sometimes erroneous or unrepresentative of the behaviour being examined. To manage this trade-off we consider the use of a Bayesian notion of surprise, which is used to perform exploration experiments whilst observations are unsurprising from the predictions that can be made and exploits when observations are surprising as they do not match the predicted response.