Exploration and Exploitation with Insufficient Resources

Chris Lovell, Gareth Jones, Klaus-Peter Zauner, Steve R. Gunn
Proceedings of the Workshop on On-line Trading of Exploration and Exploitation 2, PMLR 26:37-61, 2012.

Abstract

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.

Cite this Paper


BibTeX
@InProceedings{pmlr-v26-lovell12a, title = {Exploration and Exploitation with Insufficient Resources}, author = {Lovell, Chris and Jones, Gareth and Zauner, Klaus-Peter and Gunn, Steve R.}, booktitle = {Proceedings of the Workshop on On-line Trading of Exploration and Exploitation 2}, pages = {37--61}, year = {2012}, editor = {Glowacka, Dorota and Dorard, Louis and Shawe-Taylor, John}, volume = {26}, series = {Proceedings of Machine Learning Research}, address = {Bellevue, Washington, USA}, month = {02 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v26/lovell12a/lovell12a.pdf}, url = {https://proceedings.mlr.press/v26/lovell12a.html}, abstract = {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.} }
Endnote
%0 Conference Paper %T Exploration and Exploitation with Insufficient Resources %A Chris Lovell %A Gareth Jones %A Klaus-Peter Zauner %A Steve R. Gunn %B Proceedings of the Workshop on On-line Trading of Exploration and Exploitation 2 %C Proceedings of Machine Learning Research %D 2012 %E Dorota Glowacka %E Louis Dorard %E John Shawe-Taylor %F pmlr-v26-lovell12a %I PMLR %P 37--61 %U https://proceedings.mlr.press/v26/lovell12a.html %V 26 %X 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.
RIS
TY - CPAPER TI - Exploration and Exploitation with Insufficient Resources AU - Chris Lovell AU - Gareth Jones AU - Klaus-Peter Zauner AU - Steve R. Gunn BT - Proceedings of the Workshop on On-line Trading of Exploration and Exploitation 2 DA - 2012/05/02 ED - Dorota Glowacka ED - Louis Dorard ED - John Shawe-Taylor ID - pmlr-v26-lovell12a PB - PMLR DP - Proceedings of Machine Learning Research VL - 26 SP - 37 EP - 61 L1 - http://proceedings.mlr.press/v26/lovell12a/lovell12a.pdf UR - https://proceedings.mlr.press/v26/lovell12a.html AB - 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. ER -
APA
Lovell, C., Jones, G., Zauner, K. & Gunn, S.R.. (2012). Exploration and Exploitation with Insufficient Resources. Proceedings of the Workshop on On-line Trading of Exploration and Exploitation 2, in Proceedings of Machine Learning Research 26:37-61 Available from https://proceedings.mlr.press/v26/lovell12a.html.

Related Material