Abstract
Planning under uncertainty is a central problem in automated sequential decision making, addressed by fields ranging from AI and decision analysis to control theory and economics. This paper synthesizes these perspectives by modeling planning problems as Markov decision processes (MDPs). It identifies structural properties—such as independence in reward functions, state transitions, and feature relationships—that can be exploited to construct optimal policies efficiently. By focusing on abstraction, aggregation, and decomposition techniques, the authors demonstrate how structured, intensional representations provide the computational leverage necessary to overcome the "curse of dimensionality" in complex domains.
Introduction: Bridging the Divide
In the late 1990s, a significant gap existed between classical AI planning and decision-theoretic approaches. Classical planning excelled at managing complex, feature-rich environments but struggled with stochasticity and varying plan quality. Conversely, Operations Research utilized Markov Decision Processes (MDPs) to handle uncertainty and rewards optimally, but these models typically required an explicit (extensional) enumeration of every possible state.
As state spaces grow exponentially with the number of features—a phenomenon known as the "curse of dimensionality"—standard MDP algorithms become intractable for real-world applications. A robot with just 20 binary sensors, for instance, faces over a million states. This paper argues that the solution lies in Structural Assumptions: by exploiting the inherent structure of a problem, we can solve complex MDPs without ever enumerating the full state space.
Modeling Sequential Decisions
The core of the decision-theoretic approach is the MDP, which models an agent's interactions with an environment over time. A critical assumption in many tractable models is the Markov property, which posits that the future state depends only on the current state and action, not the history of previous states.
While a general process might have complex dependencies across all previous time steps (as shown in Figure 2a), the Markov assumption (Figure 2b and 2c) simplifies the transition model significantly. The authors extend this by proposing Factored MDPs, where the world is described by a set of variables rather than atomic states. This allows the use of 2-Stage Temporal Bayesian Networks (2TBNs) to represent actions, capturing the reality that most actions only affect a small subset of the world's features.
Illustrative Example: The Robot Delivery Domain
To demonstrate these concepts, the authors frequently reference a "Robot Coffee/Mail Delivery" example. This domain involves a robot navigating a structured environment to perform various tasks, such as delivering mail or fetching coffee.
In this environment (Figure 1), the state is defined by features like the robot's location, whether it has mail, and whether the office occupant wants coffee. A "flat" MDP would treat every combination of these features as a unique state. However, a structured approach recognizes that the action "Move to Hallway" only changes the Location variable, leaving HasMail or WantsCoffee unaffected. This local influence is the key to computational leverage.
Computational Leverage through Abstraction
The paper surveys several techniques to exploit this structure:
- Structured Policy/Value Iteration: Instead of updating a massive table of values, algorithms can operate directly on Decision Trees or Algebraic Decision Diagrams (ADDs). These representations automatically group together states that share the same value or optimal action, effectively partitioning the state space into "abstract states."
- Reachability Analysis: AI search techniques can identify states that are impossible to reach from the initial configuration, allowing the solver to prune irrelevant parts of the state space.
- Decomposition: Large problems can often be broken into smaller, independent sub-problems that can be solved in parallel or sequence.
Key Results and Impact
The synthesis provided in this paper demonstrated that structured representations could lead to exponential compression. For example, using ADDs in the SPUDD algorithm allowed researchers to solve a process-planning domain with 1.7 million states by identifying that there were only 178 distinct values across the entire space. In many structured domains, these methods proved to be orders of magnitude faster than traditional matrix-based MDP solvers.
Limitations and Future Directions
Despite the advantages, the authors note several challenges:
- Overhead: Maintaining and simplifying structured representations (like decision trees) incurs its own computational cost.
- POMDPs: Extending these techniques to Partially Observable MDPs remains difficult, as the "leverage" is harder to find when the state is not fully known.
- Automatic Discovery: There is a continuing need for methods that can automatically discover the best abstractions or decompositions without human intervention.
- Continuous Domains: Most of the discussed techniques focus on discrete variables; extending these to continuous state and action spaces remains a significant research frontier.
Summary updated