查看文章

The Relational Dynamic Influence Diagram Language (RDDL) is a uniform language where states, actions, and observations (whether discrete or continuous) are parameterized variables and the evolution of a fully or partially observed (stochastic) process is specified via (stochastic) functions over next state variables conditioned on current state and action variables (nb, concurrency is allowed). Parameterized variables are simply templates for ground variables that can be obtained when given a particular problem instance defining possible domain objects. Semantically, RDDL is simply a dynamic Bayes net (DBN)[1](with potentially many intermediate layers) extended with a simple influence diagram (ID)[2] utility node representing immediate reward. An objective function specifies how these immediate rewards should be optimized over time for optimal control. For a ground instance, RDDL is just a factored MDP (or POMDP, if partially observed).