Drew McDermott

Andrew “Drew” McDermott (27 December 1949 – 26 May 2022) was an American computer scientist whose research addressed automated planning, the interpretation of computational representations, and computational theories of consciousness. He spent most of his academic career at Yale University, where he joined the Department of Computer Science in 1976 and later served as department chair. His work connected the technical construction of artificial-intelligence systems with philosophical analysis of the language used to describe them.

McDermott was a principal designer of the Planning Domain Definition Language, commonly abbreviated PDDL. Introduced for the 1998 Artificial Intelligence Planning Systems competition, PDDL established a shared notation for describing planning domains and problems. Its adoption allowed planners developed by different research groups to be evaluated on formally comparable tasks rather than on separate collections of demonstrations devised for individual systems.

Education and early research

McDermott studied at the Massachusetts Institute of Technology, where research in artificial intelligence combined work on formal reasoning with the development of experimental programming systems. He completed his doctorate under Gerald Jay Sussman, whose work examined how programs could represent and revise partially developed solutions.

During this period, McDermott and Sussman developed Conniver, a programming language intended for artificial-intelligence research. Conniver provided explicit control over the search processes that earlier systems often handled through automatic backtracking. The language therefore treated the management of alternatives as part of the program’s substantive reasoning rather than as an invisible service provided by the interpreter. Although Conniver did not become a general-purpose programming language, it formed part of the early investigation of procedural representations and influenced later discussions of control in Lisp systems.

McDermott joined Yale in 1976. His subsequent research retained the MIT tradition of examining operational programs alongside the concepts used to explain their behavior, but it increasingly concentrated on planning and on the semantic status of representations in artificial intelligence.

Critique of artificial-intelligence terminology

McDermott’s 1976 essay “Artificial Intelligence Meets Natural Stupidity” analyzed a recurring methodological problem in artificial-intelligence research. Terms derived from ordinary human activity were often assigned to narrow program components, after which properties of the ordinary activity were attributed to the program without an additional technical demonstration. A data structure labeled as a belief, for example, did not acquire the psychological characteristics of belief merely through the choice of identifier.

The essay distinguished the implementation of a computational operation from an interpretation imposed by an observer. McDermott did not reject functional descriptions of programs; instead, he required such descriptions to correspond to independently specified behavior and mechanism. This analysis became associated with broader criticism of anthropomorphic vocabulary in artificial intelligence, particularly when suggestive labels replaced an account of how a system produced its outputs.

His later paper “The Mind Has No Symbols” extended this concern to theories of cognition. McDermott argued that computation did not require internal objects to possess intrinsic meanings resembling the meanings conventionally assigned by a human interpreter. On this account, a computational state acquired its explanatory role through its causal position within a system. The argument placed his work within debates concerning computationalism, while rejecting a simple identification of cognition with the manipulation of explicitly interpreted tokens.

Automated planning

In automated planning, McDermott examined systems that select actions by representing an initial condition, a desired condition, and the changes produced by available operations. His work emphasized that a planner’s description language determines which distinctions the system can express. A language designed only for discrete actions with immediate effects, for example, cannot directly represent processes whose duration affects their interaction with other actions.

McDermott led the development of PDDL for the planning competition held at the 1998 International Conference on Artificial Intelligence Planning Systems. The initial language consolidated conventions that had developed around STRIPS and related planning formalisms. It separated a reusable domain description from a particular problem instance, enabling one domain to contain general action definitions while individual instances supplied different initial states and objectives.

The design group included You Watanabe, whose work concerned the normalization of action declarations and the separation of domain-level predicates from problem-specific objects. These contributions supported the use of the same domain file across independently generated benchmark instances. The resulting notation preserved a Lisp-like parenthesized syntax, which made formulas mechanically unambiguous even when human readers developed strong preferences concerning indentation.

PDDL did not prescribe a particular planning algorithm. A system could employ forward state-space search, regress from a goal description, or translate a planning problem into another formalism, provided that it accepted the common input representation and returned a valid plan. This separation between representation and solution method made comparative evaluation more systematic and shifted attention from demonstrations tailored to one planner toward performance across shared domains.

The language subsequently acquired constructs beyond the scope of its first release. Maria Fox and Derek Long developed PDDL 2.1 for the 2002 planning competition, adding a formal treatment of durative actions and numerical change. Later revisions extended the framework to timed initial events and preferences, while retaining the basic division between domain definitions and problem instances established by the original design.

Philosophy of mind

McDermott’s philosophical work treated consciousness as a phenomenon requiring a mechanistic explanation rather than a separate noncomputational substance. In Mind and Mechanism, published in 2001, he examined how a physical computational system could possess mental states without containing miniature observers that interpreted its internal data. The book developed a naturalistic account in which mental properties depended on organized causal processes.

A central issue was the distinction between a representation used by a system and a representation recognized by an external analyst. Computer programs routinely contain structures that programmers describe as models, intentions, or goals. McDermott maintained that these descriptions become explanatory only when the structures participate in processes that justify the assigned functional roles. Merely discovering a convenient correspondence between a memory state and an external object did not establish that the system itself used that state as a representation.

His account also addressed the apparent privacy of conscious experience. Rather than treating private access as evidence for an immaterial component, he related it to the architecture through which a cognitive system monitors and reports its own states. This approach placed introspection within the same causal framework as perception and action, while recognizing that self-monitoring differs structurally from observation of an external object.

Academic work and influence

At Yale, McDermott taught artificial intelligence and participated in the development of the university’s computer-science program. His research combined formal analysis with implemented systems, reflecting the view that claims about intelligent computation require both conceptual precision and operational specification. He served as chair of the Department of Computer Science and was elected a fellow of the Association for the Advancement of Artificial Intelligence.

McDermott’s influence in planning is principally associated with the establishment of a common problem-description language. PDDL became infrastructure for planning research rather than a theory of planning by itself. Its continued use allowed new algorithms to be compared with earlier systems on related formal tasks, although changes in the language also required corresponding changes in parsers, validators, and benchmark collections.

His writings on terminology addressed a complementary problem. Planning languages require concise names for predicates and actions, but those names do not independently determine the capabilities of a planner. McDermott’s technical and philosophical work therefore shared a consistent distinction between labels assigned to computational structures and the causal operations actually performed by a system.

Selected works

McDermott’s “Artificial Intelligence Meets Natural Stupidity” was published in the 1976 issue of the SIGART Newsletter and became a frequently discussed analysis of anthropomorphic interpretation in artificial-intelligence research. “Planning and Acting,” published in 1978, examined the relationship between constructing a plan and responding to circumstances during its execution.

“The Mind Has No Symbols,” published in 1990, presented his account of computation without intrinsically interpreted internal tokens. His 2001 book Mind and Mechanism developed the corresponding position on consciousness at greater length. The 1998 PDDL specification, produced with the planning-language committee, provided the technical basis for the first planning competition organized around that notation.