When would you reach for it?
Scheduling, e.g., course scheduling - allocating rooms, professors, time slots while satisfying constraints; Product configuration systems - helping customers select compatible options for complex products; Genealogical research - querying family relationships and ancestry; Static analysis tools for code - finding bugs or verifying properties without execution; Medical diagnosis systems - inferring conditions from symptoms based on medical knowledge; Travel planning - finding optimal routes with multiple constraints; Legal reasoning systems - determining applicability of laws to specific cases; Natural language interfaces - parsing questions and generating appropriate database queries; Hardware verification - proving correctness of circuit designs; Puzzle solvers - Sudoku, crosswords, logic puzzles;
Basically anything that excels when declarative specification of relationships is more natural than imperative algorithms.
This all makes perfect sense. The gap I usually have - and I admit its probably something of a skill issue, I have relatively little formal CS background - is how these abstract declarations of rules are integrated into a product. The example code in projects like this is usually pretty dense and intangible.
Does anyone have good examples of open source codebases or reading material in this area? Lets imagine I have a set of complex business rules about the way a product can be configured and I want to use a logic programming language to enforce them, called from a web interface based on config data stored in a traditional relational database. Is that... a misunderstanding of how these things are to be used?
I've love a good book about how to bring tools and techniques for logical correctness into a Rails ecosystem... or similar language/framework for app dev. I love the promises many of logic languages make but can't rewrite existing applications in them wholesale and it seems like they're a poor fit for that anyways. How are people blending these worlds at large enterprises? Maybe the answer is that nobody really is yet, and thats what makes things like Clolog + Clojure so exciting?
FWIW I have no formal CS background whatsoever (maybe you shouldn't even listen to me on this matter), but if you want to gain some understanding of rule engines, you probably shouldn't start with core.logic or clolog (I have not looked into this project myself just yet, so my assumptions might be completely misleading) - core.logic imo good for complex constrains and relationships, it's based on miniKanren, but has quite steep learning curve, and not sure if it's worth the effort (as a starting point).
oakes/odoyle-rules is a forward-chaining rules engine with a more straightforward approach - for someone already familiar with Clojure, it should be fun to try out. Then maybe check out Clara Rules, if I'm not mistaken the lib is specifically designed for business rules processing. For understanding the theoretical pieces, you probably want to look into forward vs. backward chaining rule systems; pattern matching used in rules engines; understanding how to model domain rules declaratively; Rete algorithm (odoyle lib explains it and iirc links to the paper).
Great comments, thank you for taking the time to mentor a few interested strangers.
Awww, thank you. As a Clojurian, I aspire to be kind and helpful. The Clojure community is renowned for its friendly, coaching-oriented attitude. Anyone with questions (even those unrelated to Clojure) should absolutely reach out to Clojuristas — just go to http://clojurians.net and don't be shy.
I'm forever thankful for Clojure for reigniting my passion for programming, but particularly, I'm indebted to the many individuals in the Clojure community. Whenever I pose a question expecting just straightforward guidance or documentation links, I consistently receive profound, thought-provoking answers that surpass my expectations. Virtually every discussion I initiate with them ends up being incredibly educational, teaching me far more than I initially sought. I can confidently admit - yes, learning Clojure has made a better programmer out of me, but most importantly, it made me a better person.
There are lots of good Prolog books available, many from the 1980s now released freely in PDF form. I would suggest starting there. IMO, Prolog is definitely worth understanding, but I personally tend to desire a Prolog available as an embedded resource in another language. So, Clolog is great for Clojure programmers, of which I am one, for instance. I find that Prolog excels for certain types of programming, but while it’s Turing complete, and thus capable of doing anything any other language can do, it can get cumbersome for other types of programming. So, I like to think of logic programming similar to the way that I think of a SAT solver, something that I call out to in order to solve a problem, but then it returns the answer to the main program which presents that answer to the user using the host language.
The classic books can be a bit rough on the newbie, I'd recommend slogging through Triska's site and trying stuff out in Scryer, SWI or Tau first.
Well the bigger question is how big does the system have to be to warrant breaking out a new technique, much less adding a new runtime or other large dependency.
Now, I have no direct experience with any of the common logical programming systems. I have familiarity.
But anytime I came upon anything that might justify such a system, the need just didn’t seem to justify it.
Talking less than 100 rules. Most likely less than a couple dozen. Stacking some IFs and a bit of math, strategically grouped in a couple aptly named wrapper methods to help reduce the cognitive load, and it’s all worked pretty well.
And, granted, if I had solid experience using these systems, onboarding cost would be lower.
When have you found it to be worth cutting over?
I had the (self-inflicted) problem of modeling systems of linear equations in C#. `x` is a Sym, `x+2` is an Expr, `[2x,3y]` is a TermVector, etc. I wanted the comforts of NumPy, so adding an Int to a SymVector should make an ExprVector by broadcast, you should be able to multiply two IntMatrixes together but not two SymMatrixes (since that's not linear), etc. It would have been a lot of wrapper code to write.
Instead, I implemented a minimal set of primitives, and wrote a set of derivation rules (e.g. "if you have X+Y, and Y supports negation, you can derive X-Y by X+(-Y)"), and constraints (operator overloads mustn't have ambiguous signatures, no cycles allowed in the call tree), and set up a code generator.
250 lines of Prolog, plus another 250 of ASP (a dialect of Prolog), and I had a code synthesizer.
it was one of the most magical experiences of my entire career. I'd write an optimized version of a function, rerun synthesis and it would use it everywhere it could. I'd add new types and operators and it'd instantly plumb them through. seeing code synthesis dance for you feels amazingly liberating. it's like the opposite of technical debt.
For a simple problem, the equivalent of one customer demand, something like n-queens, 'placements can't conflict', the difference isn't very large between Prolog and, say, Java.
https://www.metalevel.at/queens/
https://leetcode-in-java.github.io/src/main/java/g0001_0100/...
That is, if you manage to figure out your own special case rule engine rather than a nest of if:s and for:s growing more organically.
If you have ten of these, e.g. more dimensions that would result in conflicts or constraints on where placement is possible in the domain, the Java (or PHP or JavaScript or whatever) solution is likely to turn out rather inscrutable. At least that's my experience in ERP and CRM-adjacent systems where I've spent considerable time figuring out and consolidating many years of piecemeal additions of constraint threading in things like planning and booking tasks and the like.
Sometimes I've scratched up algebraic expressions or standalone Prolog implementations to suss out what simpler code ought to look like.
Absolutely valid and befitting point - adding complexity without clear benefits never should be justified. Most (business) applications have ruleset logic for specific problems not exceeding a dozens of rules - regular code often works fine.
Logic systems tend to show the value when rules become complex with many interdependencies or non-linear execution patterns emerge, or rules change frequently or need to be defined at runtime; when you need explanation tools - e.g., why was this conclusion reached?, etc.
I agree, situations for when you need to implement a logic system are not extremely common, but maybe I just have not worked in industries where they are - on top of my head I can think of: factory floor scheduling; regulatory compliance (e.g., complex tax rules); insurance systems, risk-calculation (credit approval); strategy games; retail - complex discounting; etc.
I really don't know not nearly enough about the topic - but couldn't a lot of these be handled with types, so my question is - are types at least adjacent to logical programming? (apologies in advance)
Yes, afaik there is meaningful relevance between logic programming and type systems - Curry-Howard; Dependent types; SMT solvers (related to SAT solvers mentioned in this discussion). However, constraint problems discussed (scheduling, travel planning, puzzles) would require quite verbose and complex encoding to solve with a type system and significant sophistication - Idris or Coq aren't for the faint-hearted.
Type systems would absolutely choke on trying to solve things like: Knowlege representation and querying - imagine medical diagnosis systems with thousands symptoms and conditions; Dynamic constraint solving - where constraints emerge at runtime; Exploratory search problems requiring bfs/dfs of large solution spaces - e.g., route finding with changing conditions; NLP tasks - grammar parsing and understanding; etc.
Same reason you’d reach for SQL when querying relations - a good enough tool designed for the job.
The problem has always been getting facts into the prolog system. I’ve been looking for a prolog which is as easy to embed in eg Python or node as a Postgres client and… crickets.
I dunno which postgres client you're thinking of but:
https://github.com/tau-prolog/tau-prolog
https://pyswip.org/ https://www.swi-prolog.org/packages/mqi/prologmqi.html
Unfortunately the tau site's certificate seems to have lapsed sometime in the last week or so, but I swear it's actually very good.
Prolog has very little syntax, usually it's not particularly hard to serialise your data structure directly into Prolog code and consult it. If your data is streaming you might have to resort to assert/asserta/assertz/retract, depending on your use case.
log("2024-10-22", "09:09:23", "Mozilla", "/login").
log("2024-10-23", "09:09:24", "Safari", "/dash").
log("2024-10-24", "09:09:25", "Chrome", "/user").
log("2024-10-25", "09:09:26", "Brave", "/login").
If you already keep all your logs in analytics databases the example isn't very good, but it ought to be easy to see how this trivial technique can be applied elsewhere.