Reinforcement Learning by Policy Search

One objective of artificial intelligence is to model the behavior of an intelligent agent interacting with its environment. The environment's transformations can be modeled as a Markov chain, whose state is partially observable to the agent and affected by its actions; such processes are known as partially observable Markov decision processes (POMDPs). While the environment's dynamics are assumed to obey certain rules, the agent does not know them and must learn. In this dissertation we focus on the agent's adaptation as captured by the reinforcement learning framework. This means learning a policy---a mapping of observations into actions---based on feedback from the environment. The learning can be viewed as browsing a set of policies while evaluating them by trial through interaction with the environment. The set of policies is constrained by the architecture of the agent's controller. POMDPs require a controller to have a memory. We investigate controllers with memory, including controllers with external memory, finite state controllers and distributed controllers for multi-agent systems. For these various controllers we work out the details of the algorithms which learn by ascending the gradient of expected cumulative reinforcement. Building on statistical learning theory and experiment design theory, a policy evaluation algorithm is developed for the case of experience re-use. We address the question of sufficient experience for uniform convergence of policy evaluation and obtain sample complexity bounds for various estimators. Finally, we demonstrate the performance of the proposed algorithms on several domains, the most complex of which is simulated adaptive packet routing in a telecommunication network.

Understanding Subsystems in Biology through Dimensionality Reduction, Graph Partitioning and Analytical Modeling

Biological systems exhibit rich and complex behavior through the orchestrated interplay of a large array of components. It is hypothesized that separable subsystems with some degree of functional autonomy exist; deciphering their independent behavior and functionality would greatly facilitate understanding the system as a whole. Discovering and analyzing such subsystems are hence pivotal problems in the quest to gain a quantitative understanding of complex biological systems. In this work, using approaches from machine learning, physics and graph theory, methods for the identification and analysis of such subsystems were developed. A novel methodology, based on a recent machine learning algorithm known as non-negative matrix factorization (NMF), was developed to discover such subsystems in a set of large-scale gene expression data. This set of subsystems was then used to predict functional relationships between genes, and this approach was shown to score significantly higher than conventional methods when benchmarking them against existing databases. Moreover, a mathematical treatment was developed to treat simple network subsystems based only on their topology (independent of particular parameter values). Application to a problem of experimental interest demonstrated the need for extentions to the conventional model to fully explain the experimental data. Finally, the notion of a subsystem was evaluated from a topological perspective. A number of different protein networks were examined to analyze their topological properties with respect to separability, seeking to find separable subsystems. These networks were shown to exhibit separability in a nonintuitive fashion, while the separable subsystems were of strong biological significance. It was demonstrated that the separability property found was not due to incomplete or biased data, but is likely to reflect biological structure.

Model-Based Troubleshooting of Digital Systems

This thesis describes a methodology, a representation, and an implemented program for troubleshooting digital circuit boards at roughly the level of expertise one might expect in a human novice. Existing methods for model-based troubleshooting have not scaled up to deal with complex circuits, in part because traditional circuit models do not explicitly represent aspects of the device that troubleshooters would consider important. For complex devices the model of the target device should be constructed with the goal of troubleshooting explicitly in mind. Given that methodology, the principal contributions of the thesis are ways of representing complex circuits to help make troubleshooting feasible. Temporally coarse behavior descriptions are a particularly powerful simplification. Instantiating this idea for the circuit domain produces a vocabulary for describing digital signals. The vocabulary has a level of temporal detail sufficient to make useful predictions abut the response of the circuit while it remains coarse enough to make those predictions computationally tractable. Other contributions are principles for using these representations. Although not embodied in a program, these principles are sufficiently concrete that models can be constructed manually from existing circuit descriptions such as schematics, part specifications, and state diagrams. One such principle is that if there are components with particularly likely failure modes or failure modes in which their behavior is drastically simplified, this knowledge should be incorporated into the model. Further contributions include the solution of technical problems resulting from the use of explicit temporal representations and design descriptions with tangled hierarchies.

Generative Temporal Planning with Complex Processes

Autonomous vehicles are increasingly being used in mission-critical applications, and robust methods are needed for controlling these inherently unreliable and complex systems. This thesis advocates the use of model-based programming, which allows mission designers to program autonomous missions at the level of a coach or wing commander. To support such a system, this thesis presents the Spock generative planner. To generate plans, Spock must be able to piece together vehicle commands and team tactics that have a complex behavior represented by concurrent processes. This is in contrast to traditional planners, whose operators represent simple atomic or durative actions. Spock represents operators using the RMPL language, which describes behaviors using parallel and sequential compositions of state and activity episodes. RMPL is useful for controlling mobile autonomous missions because it allows mission designers to quickly encode expressive activity models using object-oriented design methods and an intuitive set of activity combinators. Spock also is significant in that it uniformly represents operators and plan-space processes in terms of Temporal Plan Networks, which support temporal flexibility for robust plan execution. Finally, Spock is implemented as a forward progression optimal planner that walks monotonically forward through plan processes, closing any open conditions and resolving any conflicts. This thesis describes the Spock algorithm in detail, along with example problems and test results.