Influence Networks Algorithms: Policy evaluation

From DSL

This is the main algorithm for solving influence diagrams in GeNIe. It's implementation is based on the algorithm proposed by Cooper (1988). The policy evaluation algorithm solves an influence diagram by first transforming it into a Bayesian network and then finding the expected utilities of each of the decision alternatives by performing repeated inference in this network. The algorithm will result in a full set of expected utilities for all possible policies in the network. This may be a computationally intensive process for large influence diagrams. If you are not interested in the values of expected utilities, but would just like to know the optimal decision, consider using the algorithm for finding the best decision.

GeNIe does not require the user to specify the temporal order among decision nodes in influence diagrams. However, if the order is not specified by the user, and it cannot be inferred from causal considerations (please note that directed paths starting at a decision node are necessarily causal), GeNIe will assume an order arbitrarily and make it explicit by adding arcs between decisions placing an appropriate message in the console window. GeNIe does not require the user to create non-forgetting arcs in order to avoid obscuring the structure of the model. However, it behaves as if they were there, assuming their existence from the temporal order among the decision nodes.

Finally, the policy evaluation algorithm uses the default Bayesian network algorithm specified by the user. This may have an impact on both the computational performance and the accuracy of the computation.

If you are more concerned about finding what is the optimal decision at the highest level rather than the actual utilities for each decision, Finding the best policy algorithm is a better choice. It is much faster than the Policy Evaluation algorithm.