Automated Planning: Theory & Practice (The Morgan Kaufmann Series in Artificial Intelligence) 1st Edition

Until this book, possibly the only comprehensive treatment of planning has been a paper collection: Readings in Planning (Morgan Kaufmann Series in Representation and Reasoning). What these authors have done is phenominal - they've marshalled a bibliography of 565 publications into a comprehensive treatment from a common point of view. That makes it much easier to analyze different approaches to planning, as well as to see how various application domains have applied these approaches to solve real problems.The first 448 pages of the book discusses various planning approaches, from classical state-space planning including recent improvements in the STRIPS model (GraphPlan), to dealing with temporal operations and resource scheduling. They then use the readers understanding of these deterministic approaches to bridge to planning under uncertainty, which is where planning meets the "real world" of imperfect knowledge, observability or even actions having unintended effects. The next roughly 100 pages goes into application domains discussing how space applications, robotics, manufacturing, emergency evacuation and even the game of bridge has used these planning methods to give the reader better intuitions on their own domain.Finally some minority approaches such as case-based planning and plan related areas such as plan recognition are introduced briefly, leading to tutorial appendices on search (and complexity), first order logic, and model checking.I have been working on the periphery of planning research for over 25 years, including (currently) directing advanced research in adversarial planning (a topic not addressed by this book, but that's hardly surprising given the novelty of the approach ;-). This is the best overview and reference I've seen to date for this very important area.

The book is good, covers a lot and is very clear.The downside: there are some small errors and mistakes.For example, the authors define gamma: SxAxE -> 2^S as the transition function, where S is the state space, A is the set of actions, E is the set of events. Later they say that if there are no events to be considered from the outside world, then you could use E={} (empty set) -- Assumption A3, page 10. Although this is intuitively OK, it is mathematically flawed, because the cartesian product of anything with {} is {}.Planning with MDPs and specially with POMDPs deserves more attention. In particular, the very short commentary on planning with POMDPs mentions that it is not possible to solve big POMDPs. This is not true anymore; there are very good heuristics for POMDP solving currently.I think more theorems could have been presented and proved, and some advanced sections could be added to each chapter (some authors include a section with a star, for example)I also don't like the way pseudo-code is presented, but that is a matter of taste.It would also be nice if the examples in chapter 2 were fully specified. That helps a lot to understand how problems are represented.On the good side, there are LOTS of examples for each definition, and there are exercises at the end of each chapter (more exercises would be nice, actually). I also like the discussion and historical remarks at the end of chapters.This is certainly a very good book. Anyone interested in planning ought to have it (and people interested in AI will certainly benefit from it).

This book has been recommended to me by a colleague. It is a general introduction to the domain of Robotics --- one of the most acclaimed ones.

Automated Planning is a good book for those who get started out in the field of search and planning. It's a good overview of the topics that abound within the planning community.The only downside of the book is its dealing with important topics like planning graphs and markov description process is cursory, and more detail would have been nice.