Task-Level Robot Learning

We are investigating how to program robots so that they learn from experience. Our goal is to develop principled methods of learning that can improve a robot's performance of a wide range of dynamic tasks. We have developed task-level learning that successfully improves a robot's performance of two complex tasks, ball-throwing and juggling. With task- level learning, a robot practices a task, monitors its own performance, and uses that experience to adjust its task-level commands. This learning method serves to complement other approaches, such as model calibration, for improving robot performance.

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.

Preliminary Observations on Program Instability

This white paper reports emerging findings at the end of Phase I of the Lean Aircraft Initiative in the Policy focus group area. Specifically, it provides details about research on program instability. Its objective is to discuss high-level findings detailing: 1) the relative contribution of different factors to a program’s overall instability; 2) the cost impact of program instability on acquisition programs; and 3) some strategies recommended by program managers for overcoming and/or mitigating the negative effects of program instability on their programs. Because this report comes as this research is underway, this is not meant to be a definitive document on the subject. Rather, is it anticipated that this research may potentially produce a number of reports on program instability-related topics. The government managers of military acquisition programs rated annual budget or production rate changes, changes in requirements, and technical difficulties as the three top contributors, respectively, to program instability. When asked to partition actual variance in their program’s planned cost and schedule to each of these factors, it was found that the combined effects of unplanned budget and requirement changes accounted for 5.2% annual cost growth and 20% total program schedule slip. At a rate of approximately 5% annual cost growth from these factors, it is easy to see that even conservative estimates of the cost benefits to be gained from acquisition reforms and process improvements can quickly be eclipsed by the added cost associated with program instability. Program management practices involving the integration of stakeholders from throughout the value chain into the decision making process were rated the most effective at avoiding program instability. The use of advanced information technologies was rated the most effective at mitigating the negative impact of program instability.

Preliminary Observations on Program Instability

This white paper reports emerging findings at the end of Phase I of the Lean Aircraft Initiative in the Policy focus group area. Specifically, it provides details about research on program instability. Its objective is to discuss high-level findings detailing: 1) the relative contribution of different factors to a program’s overall instability; 2) the cost impact of program instability on acquisition programs; and 3) some strategies recommended by program managers for overcoming and/or mitigating the negative effects of program instability on their programs. Because this report comes as this research is underway, this is not meant to be a definitive document on the subject. Rather, is it anticipated that this research may potentially produce a number of reports on program instability-related topics. The government managers of military acquisition programs rated annual budget or production rate changes, changes in requirements, and technical difficulties as the three top contributors, respectively, to program instability. When asked to partition actual variance in their program’s planned cost and schedule to each of these factors, it was found that the combined effects of unplanned budget and requirement changes accounted for 5.2% annual cost growth and 20% total program schedule slip. At a rate of approximately 5% annual cost growth from these factors, it is easy to see that even conservative estimates of the cost benefits to be gained from acquisition reforms and process improvements can quickly be eclipsed by the added cost associated with program instability. Program management practices involving the integration of stakeholders from throughout the value chain into the decision making process were rated the most effective at avoiding program instability. The use of advanced information technologies was rated the most effective at mitigating the negative impact of program instability.

Concurrent Aggregates (CA): An Object-Oriented Language for Fine-Grained Message-Passing Machines

Fine-grained parallel machines have the potential for very high speed computation. To program massively-concurrent MIMD machines, programmers need tools for managing complexity. These tools should not restrict program concurrency. Concurrent Aggregates (CA) provides multiple-access data abstraction tools, Aggregates, which can be used to implement abstractions with virtually unlimited potential for concurrency. Such tools allow programmers to modularize programs without reducing concurrency. I describe the design, motivation, implementation and evaluation of Concurrent Aggregates. CA has been used to construct a number of application programs. Multi-access data abstractions are found to be useful in constructing highly concurrent programs.

The Role of Programming in the Formulation of Ideas

Classical mechanics is deceptively simple. It is surprisingly easy to get the right answer with fallacious reasoning or without real understanding. To address this problem we use computational techniques to communicate a deeper understanding of Classical Mechanics. Computational algorithms are used to express the methods used in the analysis of dynamical phenomena. Expressing the methods in a computer language forces them to be unambiguous and computationally effective. The task of formulating a method as a computer-executable program and debugging that program is a powerful exercise in the learning process. Also, once formalized procedurally, a mathematical idea becomes a tool that can be used directly to compute results.

The Definition and Implementation of a Computer Programming Language Based on Constraints

The constraint paradigm is a model of computation in which values are deduced whenever possible, under the limitation that deductions be local in a certain sense. One may visualize a constraint 'program' as a network of devices connected by wires. Data values may flow along the wires, and computation is performed by the devices. A device computes using only locally available information (with a few exceptions), and places newly derived values on other, locally attached wires. In this way computed values are propagated. An advantage of the constraint paradigm (not unique to it) is that a single relationship can be used in more than one direction. The connections to a device are not labelled as inputs and outputs; a device will compute with whatever values are available, and produce as many new values as it can. General theorem provers are capable of such behavior, but tend to suffer from combinatorial explosion; it is not usually useful to derive all the possible consequences of a set of hypotheses. The constraint paradigm places a certain kind of limitation on the deduction process. The limitations imposed by the constraint paradigm are not the only one possible. It is argued, however, that they are restrictive enough to forestall combinatorial explosion in many interesting computational situations, yet permissive enough to allow useful computations in practical situations. Moreover, the paradigm is intuitive: It is easy to visualize the computational effects of these particular limitations, and the paradigm is a natural way of expressing programs for certain applications, in particular relationships arising in computer-aided design. A number of implementations of constraint-based programming languages are presented. A progression of ever more powerful languages is described, complete implementations are presented and design difficulties and alternatives are discussed. The goal approached, though not quite reached, is a complete programming system which will implicitly support the constraint paradigm to the same extent that LISP, say, supports automatic storage management.

Inspection Methods in Programming

The work reported here lies in the area of overlap between artificial intelligence software engineering. As research in artificial intelligence, it is a step towards a model of problem solving in the domain of programming. In particular, this work focuses on the routine aspects of programming which involve the application of previous experience with similar programs. I call this programming by inspection. Programming is viewed here as a kind of engineering activity. Analysis and synthesis by inspection area prominent part of expert problem solving in many other engineering disciplines, such as electrical and mechanical engineering. The notion of inspections methods in programming developed in this work is motivated by similar notions in other areas of engineering. This work is also motivated by current practical concerns in the area of software engineering. The inadequacy of current programming technology is universally recognized. Part of the solution to this problem will be to increase the level of automation in programming. I believe that the next major step in the evolution of more automated programming will be interactive systems which provide a mixture of partially automated program analysis, synthesis and verification. One such system being developed at MIT, called the programmer's apprentice, is the immediate intended application of this work. This report concentrates on the knowledge are of the programmer's apprentice, which is the form of a taxonomy of commonly used algorithms and data structures. To the extent that a programmer is able to construct and manipulate programs in terms of the forms in such a taxonomy, he may relieve himself of many details and generally raise the conceptual level of his interaction with the system, as compared with present day programming environments. Also, since it is practical to expand a great deal of effort pre-analyzing the entries in a library, the difficulty of verifying the correctness of programs constructed this way is correspondingly reduced. The feasibility of this approach is demonstrated by the design of an initial library of common techniques for manipulating symbolic data. This document also reports on the further development of a formalism called the plan calculus for specifying computations in a programming language independent manner. This formalism combines both data and control abstraction in a uniform framework that has facilities for representing multiple points of view and side effects.

Building Grounded Abstractions for Artificial Intelligence Programming

Most Artificial Intelligence (AI) work can be characterized as either ``high-level'' (e.g., logical, symbolic) or ``low-level'' (e.g., connectionist networks, behavior-based robotics). Each approach suffers from particular drawbacks. High-level AI uses abstractions that often have no relation to the way real, biological brains work. Low-level AI, on the other hand, tends to lack the powerful abstractions that are needed to express complex structures and relationships. I have tried to combine the best features of both approaches, by building a set of programming abstractions defined in terms of simple, biologically plausible components. At the ``ground level'', I define a primitive, perceptron-like computational unit. I then show how more abstract computational units may be implemented in terms of the primitive units, and show the utility of the abstract units in sample networks. The new units make it possible to build networks using concepts such as long-term memories, short-term memories, and frames. As a demonstration of these abstractions, I have implemented a simulator for ``creatures'' controlled by a network of abstract units. The creatures exist in a simple 2D world, and exhibit behaviors such as catching mobile prey and sorting colored blocks into matching boxes. This program demonstrates that it is possible to build systems that can interact effectively with a dynamic physical environment, yet use symbolic representations to control aspects of their behavior.

On the Convergence of Stochastic Iterative Dynamic Programming Algorithms

Recent developments in the area of reinforcement learning have yielded a number of new algorithms for the prediction and control of Markovian environments. These algorithms, including the TD(lambda) algorithm of Sutton (1988) and the Q-learning algorithm of Watkins (1989), can be motivated heuristically as approximations to dynamic programming (DP). In this paper we provide a rigorous proof of convergence of these DP-based learning algorithms by relating them to the powerful techniques of stochastic approximation theory via a new convergence theorem. The theorem establishes a general class of convergent algorithms to which both TD(lambda) and Q-learning belong.