decisionmaking: the errata tracker for a decision making textbook

A textbook, and a repository that only tracks its corrections

Algorithms for Decision Making is written by Mykel Kochenderfer, Tim Wheeler and Kyle Wray, and the book provides a broad introduction to algorithms for decision making under uncertainty, covering a wide variety of topics related to decision making and introducing both the underlying mathematical problem formulations and the algorithms for solving them.

That is the whole project description, and it sits above a repository that contains exactly one file, README.md. There is no source, no build, no test directory and no licence file. The top level listing is a single entry.

So the repository is not where the book lives. It is where the corrections to the book live. The readme's opening instruction is to file issues for suggestions and comments, or to email the address listed at the bottom of the pages of the PDF, which tells you the artefact being corrected is a PDF edition and that the authors consider a printed or rendered page the unit of reference.

The book's own web page is algorithmsbook.com, and that is the place to find the book itself.

The choice to keep the corrections in a public repository rather than in a printed errata sheet or a new edition is a deliberate one, and it is the same reason publishers now do this: a book that is revised continuously in the open does not need a second edition for a sign error. The list of corrections below is the evidence for that claim.

Forty five corrections, each with a page and a reason

The errata section is headed First printing, and the entries run from page 73 to page 644 in the appendix. Each entry names a page, an equation, algorithm, exercise, figure or section, states the correction, and credits the person who reported it.

A few of them, verbatim, to show the format.

* [p. 97] Eq 5.3: Instead of equality, it should be propto (thanks to Esen Yel)
* [p. 99] Sec 5.2: In the first paragraph, change 2.4 to 2.3 (thanks to Dylan Asmar)
* [p. 108] Ex 5.4: Make edge E->D in the PDAG directed (thanks to Nikhil Raghuraman)
* [p. 121] Alg 6.2: Caption updated to reflect that algorithm returns the value of information (thanks to Griffin Holt)
* [p. 174] Fig 8.13: Changed "polynomial" to "quadratic" (thanks to Balduin)

The credits are the notable part. The corrections run to dozens of entries, and nearly every one has a name attached, with several people appearing more than once. Esen Yel, Dylan Asmar, Griffin Holt and Paul Diederichs each appear at least twice, and some entries credit a group, such as the entry about instances of the letter l that should be the script ell in an exercise, which is credited to the publisher, Asakura publishing.

The mix of contributors tells you something about the audience. The people reporting these are the people reading the book carefully enough to notice that an algorithm returns an action index rather than an action, that a caption should say best response rather than response, that a line plot needs colours that work for colour blindness, or that a sentence in a figure caption was left in the wrong form. This is a book being used as a reference by people who work in the area, and the errata is the peer review that never quite stopped.

What the corrections reveal about the book's contents

A list of errata is unusual evidence, because it points at the parts of a book that readers actually lean on. Reading the entries in order gives a rough map of the volume.

The early chapters are probability and Bayesian modelling. Equation 4.10 is a Gaussian log likelihood over a set of observations, equation 5.3 concerns proportionality rather than equality, and there is a correction to a paragraph in section 5.2 that changes a cross reference from 2.4 to 2.3, alongside an entry noting that the Bayesian score drops the constant in the log probability term. A chapter on graphical models follows, with a correction making an edge in a PDAG directed.

The middle of the book covers sequential decision making and machine learning. Algorithm 6.2 has its caption updated to reflect that it returns the value of information, and there are corrections to a greedy algorithm, to a figure caption that changed polynomial to quadratic, and to a statement that the singular value decomposition should be X equals U S V, which tells you there is a section on principal component analysis and matrix factorisation.

Several chapters concern reinforcement learning and bandits. Exercise 15.2 changed the model creation code to model equals BanditModel with a Beta, and the example pinned a random number generator to reproduce the expected output, which is a very specific kind of correction to need. There is a note about what to do when the count is zero, a change to a posterior probability expression, and a note that a policy should be prepended with the letter pi.

Later chapters handle games and multi agent settings, with a correction changing a response policy to a best response policy in section 25.2.1, a table caption changed from Dec-POMDP to Dec-MDP, and a series of corrections to observation functions in chapters 19 to 22, including a change from a perfect observation to a more accurate one.

The appendices are not exempt. There is a sign fix in an equation in appendix D, and a note in appendix G that a letter in the text should be capitalised to be consistent with the code.

The kinds of feedback the authors want

The readme lists the forms of feedback the project is interested in, and the list is worth reading as a statement of what the authors think a textbook needs from its readers.

The six categories are errors, improvements to code, typos, areas that are confusing, critical topics that are missing, and ideas for examples or exercises. The framing is that all of these are welcome, and the request is to file issues or send an email rather than to open a pull request, which tells you the repository is not where the text is edited.

The entry on code carries a qualifier that says a lot about the intended audience. Improvements to code are wanted, and the parenthetical says especially improvements for clarity over speed. A reader is being asked to make the pseudocode clearer, not faster, which is the right request for a book whose algorithms are there to be understood and reimplemented, and it is a request that few contributors would make unprompted.

The category that is missing, and the one that makes this list valuable, is areas that are confusing. Combined with critical topics that are missing, it means the authors are treating reader confusion as a defect worth recording, not as a gap in the reader's background. That is the same instinct as the errata list itself, applied to the prose rather than to the formulas.

One entry in the list is not a typo but a substantive request, and it is easy to miss in the middle of a long page reference. On page 436, section 21.6 and the implementation of the utility function in algorithm 21.9 were updated, which is a change to a piece of code in the text, and the credit goes to a named contributor rather than to the publisher.

Where this repository is the wrong tool

Four limits, and the first is the one that decides whether this repository is useful to you at all.

The corrections apply to the first printing, and the list says so. If you own a later printing, many of these fixes may already be in the text you are reading, and if you own a different edition, the page references may not line up with anything. Before applying an erratum, check it against the printing in your hands. The list is also not versioned by anything else, since the repository publishes no releases, so there is no way to ask which corrections were known at a given date.

The second is that there is no code. The book contains algorithms and pseudocode, and a reader who wants to run them has nothing here to run. The errata mention things like a model creation line and a function implementation, but they are descriptions of text, not artefacts you can execute.

The third is the absence of any licence statement. The repository lists no licence and contains no licence file, and the work it tracks is a book with a publisher. If you want to reuse the errata, the pseudocode or the text in your own course materials or your own book, you have no stated permission to point at, and that question needs an answer from the authors rather than a reading of the repository.

The fourth is scale. There are dozens of entries for a volume of more than six hundred pages, and the fixes range from a missing minus sign in an appendix equation to a rewritten paragraph. That is a healthy rate, and it also means the list is long enough to be hard to search, so a reader in a hurry should use the page and equation references rather than reading it end to end.

Maintenance, and what the repository does not tell you

The repository is not archived, and the last push was on 2026-10-14, so the errata list is being extended. There are no GitHub releases, which for an errata log is not a problem at all, since the list is append only and the page references are the identifiers that matter.

The list is organised in the way a reader would want it. Each entry is anchored to a page, an equation number, an algorithm number, an exercise number or a section, which means you can find the correction for a specific piece of the book rather than scanning prose. The first printing heading tells you which printing the corrections apply to, and the credits tell you who found each one, which means a reader who wants to follow up on a specific point knows who to ask.

What the repository does not tell you is almost everything else. There is no description of the book's structure, no table of contents, no chapter list, no information about the exercises or the code listings, no publication details, and no statement of licence. The description line in the repository metadata is four words long, and the readme repeats a paragraph of it.

For a project whose entire purpose is to accumulate small corrections in public, that is a defensible amount of ceremony. It also means that a reader arriving at the repository with no copy of the book in hand will find the errata list and little else, which is the shape of the artefact: it is a companion to a text, not a substitute for it. The book webpage at algorithmsbook.com is where the text and the surrounding material live, and the errors, the typos, the confusing passages and the missing topics all belong in the issue tracker of this repository.

Editorial conclusion

decisionmaking is worth following if you own or teach from Algorithms for Decision Making, because the errata are precise, credited and organised by page and equation, which is more than most textbooks manage. It is the wrong place to look for the book itself, for the code behind the examples, or for any statement about what licence the book is under, since the repository records none. Before you rely on any corrected equation from the list, check the page number against your own printing, because the list is explicitly for the first printing and later printings may already incorporate the fixes.

Frequently asked questions

Official sources

1. algorithmsbooks/decisionmaking on GitHub
2. Issues
3. README