Adaptive robust optimization
Author: Woo Soo Choe (ChE 345 Spring 2015)
Steward: Dajun Yue, Fengqi You
Traditionally, robust optimization has solved problems based on static decisions which are predetermined by the decision makers. Once the decisions were made, the problem was solved and whenever a new uncertainty was realized, the uncertainty was incorporated to the original problem and the entire problem was solved again to account for the uncertainty. Generally, robust optimization problem is formulated as follows.
In the equation
Unfortunately, most robust optimization problems for real life applications require multiple stages to account for uncertainties and traditional static robust has shown limitations. In order to improve the pre-existing technique, adaptive robust optimization was studied and advances in the field was made to address the problems which could not be easily handled with previous methods.
Adaptive robust optimization implements different techniques to improve on the original static robust optimization by incorporating multiple stages of decision into the algorithm. Currently, in order to minimize the complexity of algorithm, most of the studies on adaptive robust optimization have focused on two-stage problems. Generally, adaptive robust optimization has the basic set up as follows.
In the preceding expression, the p's are the parameters of the objective function and constraints. A noteworthy feature of the parameters is that they have aspects of ambiguity in them. The vector y is the first stage decision made prior to the realization of ambiguity. First stage decisions are based on a finite number of set of possible scenarios with unknown probability distribution. Furthermore, first stage decisions are chosen in such a way that no assumption is required for the decision parameter. This reduces the need for the decision makers to come up with any ambiguous parameters. Such use of first-stage decision variables is called robust decision setting. The vector x represents the second stage decision variable after the amibguity is realized. The overall algorithm sequentially solves and updates a relaxation problem unitl both feasibility and optimum conditions of the overall problems are met. The method also makes use of enumerative search in order to reach the answer. Adaptive robust maybe used make a minmax regret decision. In other words, when a firm needs to make a decision involving risk and uncertainty, it may utilize adaptive robust optimization to take the path that leads to minimum loss when the worst case scenario takes place.
In the early 1970's, Soyster proposed an adaptive robust optimization method which utilizes linear optimization model to come up with a solution that work any value from an interval. This method was able to account for the uncertainty and yield an answer that met the specification of the algorithm. This method often yielded overly-conservative result and Ben-Tal, Nemirovski and others resolved the issue by representing the uncertainty sets as ellipsoids and devising an algorithm that solves convex optimization problems when the data has uncertainty. Even with the improvements, these techniques have drawbacks in regards to real-life application, because these approaches involve conic quadratic problems, these methods cannot be easily implemented to discrete cases. Given vast variety of algorithm in the field of adaptive robust optimization, only scenario relaxation algorithm for finite scenario based min-max regret will be studied in this wiki page.
In order to carry out adaptive robust optimization using scenario based relaxation algorithm, the number of the possible scenarios is reduced based on two criteria. Under the first criterion, the number of scenarios are reduced based on whether the solution is feasible for all possible scenarios. Under the second criterion, the number of the scenarios is reduced so that the scenarios would result in the optimal robust solution. The algorithm works in a way that the first subset of possible scenarios are used to solve the following equation.
Then, the second subset of possible scenarios are used to calculate the regrets at each iteration for all scenarios which were previously ruled out in the possible scenario categorization. If the optimal objective value obtained from running the first subset is greater than or equal to the calculated regret from optimality, the optimal condition is met and the algorithm terminates. If not, previously not included scenarios would be included and then the iteration will repeat until the optimum solution is found.