[4] In that work, he coined the term "task network" to refer to hierarchies of abstract and concrete actions and policies.
McDermott did seminal work in non-monotonic logic in the early 1980s and was an advocate for the "logicist" methodology in AI, defined as formalizing knowledge and reasoning in terms of deduction and quasideduction.
The critique was based partly on a previous paper (with Steve Hanks) pointing out a flaw with all known approaches to nonmonotonic temporal reasoning, embodied in what is now called the Yale shooting problem.
Although new approaches have since been found, McDermott turned to other areas of AI, such as vision and robotics, and began working on automated planning again.
In 1996, McDermott (and Hector Geffner and Blai Bonet independently) discovered "estimated-regression planning," based on the idea of heuristic search with an estimator derived from a simplified domain model by reasoning backward ("regression") from the goal.