Will Taylor

AutoClass: a Bayesian classification system

This paper describes AutoClass II, a program for automatically discovering (inducing) classes from a database, based on a Bayesian statistical technique which automatically determines the most probable number of classes, their probabilistic descriptions, and the probability that each object is a member of each class. AutoClass has been tested on several large, real databases and has discovered previously unsuspected classes. There is no doubt that these classes represent new phenomena.

Spacecraft Autonomy Flight Experience: The DS1 Remote Agent Experiment

In May 1999 state-of-the-art autonomy technology was allowed to assume command and control of the Deep Space One spacecraft during the Remote Agent Experiment. This experiment demonstrated numerous autonomy concepts ranging from high-level goaloriented commanding to on-board planning to robust plan execution to model-based fault protection. Many lessons of value to future enhancements of spacecraft autonomy were learned in preparing for and executing this experiment. This paper describes those lessons and suggests directions of future work in this field.

Computational Complexity And Phase Transitions

It is well known that for many NP-complete problems, such as I<-Sat, I<-colorability etc., typical cases are easy to solve; so that coniputationally hard cases must be rare (assuming P # NP). This paper shows that NP-complete problems can be summarized by at least one “order parameter”, and that the hard problems occur at a critical value of such a parameter. This critical value separates two regions of characteristically different properties. For example, for K-colorability, the critical value separates overconstrained from underconstrained random graphs, and it marks the value at which the probability of a solution changes abruptly from near 0 to near 1. It is the high density of well-separated almost solutions (local minima) at this boundary that cause search algorithms to “thrash”. This boundary is a type of phase transition and we show that it is preserved under mappings between problems. We show that for some P problems either there is no phase transition or it occurs for bounded N (and so bounds the cost). These results suggest a way of deciding if a problem is in P or NP and why they are different.

Ground tools for autonomy in the 21st century

Ground tools for unmanned spacecraft are changing rapidly driven by twin innovations: advanced autonomy and ubiquitous networking. Critical issues are the delegation of low-level decision-making to software, the transparency and accountability of that software, mixed-initiative control, i.e., the ability of controllers to adjust portions of the softwares activity without disturbing other portions, and the makeup and geographic distribution of the flight control team. These innovations will enable ground controllers to manage space-based resources much more efficiently and, in the case of science missions, give principal investigators an unprecedented level of direct control. This paper explores these ideas by describing the ground tools for the Remote Agent experiment aboard the Deep Space 1 spacecraft in May of 1999. The experiment demonstrated autonomous control capabilities including goal-oriented commanding, on-board planning, robust plan execution and model-based fault protection. We then speculate on the effect of these technologies on the future of spacecraft ground control.

PLASMA: a constraint based planning architecture

Until recently, planning research focused on solving problems of feasibility using models consisting of causal rules. Propositional logic is sufficient for representing such rules. However, many planning problems also contain time and resource constraints. It is often impractical to represent such planning domains with propositions. Large time horizons and possible resource states lead to enormous domain representations. Propositional representations can often obscure information that is useful during search. Finally, propositional representations can make it difficult for human modelers to express the domain in a convenient and natural way. The Constraint-based Planning paradigm employs constraints as the building blocks of both planning domain rules and plans. The building blocks of such plans are intervals of time over which some state holds or an action occurs in a plan. Each interval is represented by variables describing its properties (e.g. start, end, duration). At each step of the planning process, a mapping is maintained between the plan under construction and an underlying constraint network. As actions are added to the plan, constraints are posted on variables representing the action, its preconditions and its effects (a generalization of causal links). The domain rules contain directives for adding new intervals and for posting constraints over the variables on those intervals as plans are modified. Employing constraints in rules makes it easy to represent disjunctive preconditions, conditional effects, and mutual exclusions directly. The semantics of this mapping ensure that logical inference (e.g. propagation) on the constraint network can be used directly by search engines operating on plans.