Improving Sequential Decisions – Efficiently Accounting for Future Learning

Abstract In sequential field development planning, past decisions not only directly affect the maximum achievable expected NPV but also influence the future information that can be used to reduce geological uncertainty. To act optimally, when choosing actions, we must also take into account the opportunities to improve the optimal strategy by reducing future uncertainty. In most applications, however, the effect of future information on the optimal decisions is ignored because it would be computationally intractable to update the reservoir model and re-optimize to account for all possible outcomes of future observations. To efficiently make optimal decisions while considering future possibilities for learning through actions, we developed a flexible workflow built on the key-feature-based value of information (VOI) analysis, which is obtained by identifying key reservoir features for optimization problems and key observations for improving future decisions. Instead of considering future information from all remaining actions, we only consider the important information from key actions to reduce the uncertainty with the largest influence on the optimal strategy – that which would be most helpful in improving future decisions. The efficiency of the method results from the focus on the use of key observations to reduce key uncertainty, rather than using all observations to reduce all uncertainties. In this work, we built supervised-learning algorithms to identify the optimal combination of observations for reducing key uncertainty and simultaneously to estimate the information’s reliability. This allows automatic detection of key observations and direct computation of the posterior probability distribution of key uncertainty based on Bayes’ rule, avoiding the need for full history matching to re-estimate the uncertainty. Moreover, the entire key observation space is divided into a limited number of disjointed subspaces, such that observations located in the same subspace have almost the same prediction precision for key uncertainty reduction. It is then only necessary to update the reservoir model for each subspace instead of for all distinct sets of observations. Our methods are illustrated by the application of the drilling-order problem in a synthetic field model, for which the drilling sequence of wells is an important contributor to the reservoir’s profitability and for which the optimal solution changes significantly with key reservoir features. Results show that using such a simplified VOI analysis based on key reservoir features and key observations can efficiently improve the expected outcome of an optimal strategy with very little performance loss. Although the key actions provide important information to reduction of key uncertainty and improve the optimal strategy, taking key action rather than the initial optimal decision for the current uncertainty state is not always worthwhile even if the information is obtained without explicit cost. Since there may be an indirect cost of information caused by taking an action that appears to be sub-optimal based on past information, it is necessary to consider both the possibility of key uncertainty reduction and the possibility of high expected NPV to determine whether it is worth taking the action to improve future decisions.