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.

Electrical Design: A Problem for Artificial Intelligence Research

This report outlines the problem of intelligent failure recovery in a problem-solver for electrical design. We want our problem solver to learn as much as it can from its mistakes. Thus we cast the engineering design process on terms of Problem Solving by Debugging Almost-Right Plans, a paradigm for automatic problem solving based on the belief that creation and removal of "bugs" is an unavoidable part of the process of solving a complex problem. The process of localization and removal of bugs called for by the PSBDARP theory requires an approach to engineering analysis in which every result has a justification which describes the exact set of assumptions it depends upon. We have developed a program based on Analysis by Propagation of Constraints which can explain the basis of its deductions. In addition to being useful to a PSBDARP designer, these justifications are used in Dependency-Directed Backtracking to limit the combinatorial search in the analysis routines. Although the research we will describe is explicitly about electrical circuits, we believe that similar principles and methods are employed by other kinds of engineers, including computer programmers.

Automatic Analysis and Synthesis of Controllers for Dynamical Systems Based On P

I present a novel design methodology for the synthesis of automatic controllers, together with a computational environment---the Control Engineer's Workbench---integrating a suite of programs that automatically analyze and design controllers for high-performance, global control of nonlinear systems. This work demonstrates that difficult control synthesis tasks can be automated, using programs that actively exploit and efficiently represent knowledge of nonlinear dynamics and phase space and effectively use the representation to guide and perform the control design. The Control Engineer's Workbench combines powerful numerical and symbolic computations with artificial intelligence reasoning techniques. As a demonstration, the Workbench automatically designed a high-quality maglev controller that outperforms a previous linear design by a factor of 20.

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.

Embodiment and Manipulation Learning Process for a Humanoid Hand

Babies are born with simple manipulation capabilities such as reflexes to perceived stimuli. Initial discoveries by babies are accidental until they become coordinated and curious enough to actively investigate their surroundings. This thesis explores the development of such primitive learning systems using an embodied light-weight hand with three fingers and a thumb. It is self-contained having four motors and 36 exteroceptor and proprioceptor sensors controlled by an on-palm microcontroller. Primitive manipulation is learned from sensory inputs using competitive learning, back-propagation algorithm and reinforcement learning strategies. This hand will be used for a humanoid being developed at the MIT Artificial Intelligence Laboratory.

Automated Reasoning About Classical Mechanics

In recent years, researchers in artificial intelligence have become interested in replicating human physical reasoning talents in computers. One of the most important skills in this area is predicting how physical systems will behave. This thesis discusses an implemented program that generates algebraic descriptions of how systems of rigid bodies evolve over time. Discussion about the design of this program identifies a physical reasoning paradigm and knowledge representation approach based on mathematical model construction and algebraic reasoning. This paradigm offers several advantages over methods that have become popular in the field, and seems promising for reasoning about a wide variety of classical mechanics problems.

A Trainable Object Detection System: Car Detection in Static Images

This paper describes a general, trainable architecture for object detection that has previously been applied to face and peoplesdetection with a new application to car detection in static images. Our technique is a learning based approach that uses a set of labeled training data from which an implicit model of an object class -- here, cars -- is learned. Instead of pixel representations that may be noisy and therefore not provide a compact representation for learning, our training images are transformed from pixel space to that of Haar wavelets that respond to local, oriented, multiscale intensity differences. These feature vectors are then used to train a support vector machine classifier. The detection of cars in images is an important step in applications such as traffic monitoring, driver assistance systems, and surveillance, among others. We show several examples of car detection on out-of-sample images and show an ROC curve that highlights the performance of our system.

Factorial Hidden Markov Models

We present a framework for learning in hidden Markov models with distributed state representations. Within this framework, we derive a learning algorithm based on the Expectation--Maximization (EM) procedure for maximum likelihood estimation. Analogous to the standard Baum-Welch update rules, the M-step of our algorithm is exact and can be solved analytically. However, due to the combinatorial nature of the hidden state representation, the exact E-step is intractable. A simple and tractable mean field approximation is derived. Empirical results on a set of problems suggest that both the mean field approximation and Gibbs sampling are viable alternatives to the computationally expensive exact algorithm.

Cooperative Physics of Fly Swarms: An Emergent Behavior

We have simulated the behavior of several artificial flies, interacting visually with each other. Each fly is described by a simple tracking system (Poggio and Reichardt, 1973; Land and Collett, 1974) which summarizes behavioral experiments in which individual flies fixate a target. Our main finding is that the interaction of theses implemodules gives rise to a variety of relatively complex behaviors. In particular, we observe a swarm-like behavior of a group of many artificial flies for certain reasonable ranges of our tracking system parameters.

Object Detection in Images by Components

In this paper we present a component based person detection system that is capable of detecting frontal, rear and near side views of people, and partially occluded persons in cluttered scenes. The framework that is described here for people is easily applied to other objects as well. The motivation for developing a component based approach is two fold: first, to enhance the performance of person detection systems on frontal and rear views of people and second, to develop a framework that directly addresses the problem of detecting people who are partially occluded or whose body parts blend in with the background. The data classification is handled by several support vector machine classifiers arranged in two layers. This architecture is known as Adaptive Combination of Classifiers (ACC). The system performs very well and is capable of detecting people even when all components of a person are not found. The performance of the system is significantly better than a full body person detector designed along similar lines. This suggests that the improved performance is due to the components based approach and the ACC data classification structure.