Abstract
We consider the problem of estimating the value of a multiattribute resource, where the attributes are categorical or discrete in nature and the number of potential attribute vectors is very large. The problem arises in approximate dynamic programming when we need to estimate the value of a multiattribute resource from estimates based on Monte-Carlo simulation. These problems have been traditionally solved using aggregation, but choosing the right level of aggregation requires resolving the classic tradeoff between aggregation error and sampling error. We propose a method that estimates the value of a resource at different levels of aggregation simultaneously, and then uses a weighted combination of the estimates. Using the optimal weights, which minimizes the variance of the estimate while accounting for correlations between the estimates, is computationally too expensive for practical applications. We have found that a simple inverse variance formula (adjusted for bias), which effectively assumes the estimates are independent, produces near-optimal estimates. We use the setting of two levels of aggregation to explain why this approximation works so well.
Original language | English (US) |
---|---|
Pages (from-to) | 2079-2111 |
Number of pages | 33 |
Journal | Journal of Machine Learning Research |
Volume | 9 |
State | Published - Oct 2008 |
All Science Journal Classification (ASJC) codes
- Software
- Artificial Intelligence
- Control and Systems Engineering
- Statistics and Probability
Keywords
- Adaptive learning
- Approximate dynamic programming
- Hierarchical statistics
- Mixture models
- Multiattribute resources