Skip to content
FlexianaPublic

About

Operations research in higher-order functions: a dependency-free Clojure library for optimization, decision making, and validation. Companion to the book 'Applied Higher Order Functions'.

Resources

Stars

20 stars

Watchers

0 watching

Forks

Repository files navigation

orhof — Operations Research in Higher-Order Functions

A dependency-free Clojure library showing that a large body of operations-research algorithms — continuous and discrete optimization, sequential decision making, and validation — are compositions of a small set of higher-order function (HOF) patterns.

It is the companion library to the book Applied Higher Order Functions: A Functional Architecture of Optimization, Decision Making, and Validation (Jiri Knesl). Every code listing in the book runs against this library.

  • 294 public functions across 34 namespaces
  • Zero dependencies beyond Clojure 1.12
  • On Clojars: com.flexiana/orhof {:mvn/version "1.0.0"}
  • Three domains: optimization · Markov decision processes / RL · temporal-logic validation
  • One tree-search core (search, explore, hylo) under the book's tree-search algorithms

Install

Clojars Project

Clojure CLI, deps.edn:

com.flexiana/orhof {:mvn/version "1.0.0"}

Leiningen, project.clj:

[com.flexiana/orhof "1.0.0"]

Git dependency:

com.flexiana/orhof {:git/url "https://github.com/Flexiana/orhof"
                    :git/tag "v1.0.0" :git/sha "f22436d"}

Try it in a REPL without a project:

clj -Sdeps '{:deps {com.flexiana/orhof {:mvn/version "1.0.0"}}}'

Needs Clojure 1.12 or later and nothing else.

Quick start

(require '[orhof.optimize :as opt]
         '[orhof.mdp      :as mdp]
         '[orhof.validate :as val])

;; --- Optimization: gradient descent on a paraboloid ---
(opt/gradient-descent (fn [[x y]] (+ (* x x) (* y y))) [5.0 5.0] :lr 0.1 :max-iter 500)
;; => {:x [~0.0 ~0.0] :value ~0.0 :iterations n :converged? true ...}

;; Same problem, Adam — swap one function, same framework
(opt/adam (fn [[x y]] (+ (* x x) (* y y))) [5.0 5.0] :lr 0.1 :max-iter 500)

;; --- Decision making: a multi-armed bandit via Thompson sampling ---
(def bandit (mdp/make-bandit [(fn [] (if (< (rand) 0.3) 1.0 0.0))
                              (fn [] (if (< (rand) 0.7) 1.0 0.0))]))
(mdp/thompson-sampling bandit :n-rounds 2000)   ; => {:alphas ... :betas ... :counts ...}

;; --- Validation: falsify a temporal-logic specification ---
;; see orhof.validate / orhof.val.* and docs/API.md

;; --- Tree search: unfold a space and fold it in the same pass ---
(require '[orhof.search :refer [explore search hylo breadth-first best-first unique]])

(take 5 (explore breadth-first (fn [n] [(* 2 n) (inc (* 2 n))]) 1))   ; an infinite tree
;; => (1 2 3 4 5)

((hylo (fn [[x y]] (cond-> [] (< x 16) (conj [(inc x) y]) (< y 16) (conj [x (inc y)])))
       (fn [_ counts] (if (seq counts) (reduce + counts) 1N))
       :memo identity)                                               ; dynamic programming
 [0 0])
;; => 601080390N   paths across a 16x16 grid

Namespaces

Namespace What it provides
orhof.core HOF primitives — iterative refinement, function transformers, pipeline composition, convergence predicates — plus vector math
orhof.math linear algebra, calculus, and probability utilities
orhof.diff differentiation as a HOF: gradient, hessian, finite differences, autodiff
orhof.optimize, orhof.opt.* the generic descent framework, Newton/quasi-Newton, Adam & friends, genetic algorithms, PSO, simulated annealing, constrained (penalty / barrier / augmented Lagrangian), linear/quadratic, surrogates, multi-objective
orhof.mdp, orhof.mdp.* value & policy iteration, Q-learning, MCTS, bandits, belief filters (Kalman/particle), games/equilibria, inference
orhof.validate, orhof.val.* temporal-logic specs (always, eventually, until), robustness semantics, falsification, reachability, sampling, counterfactual explanation
orhof.search the tree-search pattern as composable HOFs: strategies (depth-first, breadth-first, best-first, unique), search, explore, hylo, hylo-eager, converge
orhof.search.* algorithms built on it: shortest paths (graph), minimax, alpha-beta, expectiminimax, MCTS (games), forward search, branch and bound, sparse sampling, belief-space planning (mdp), adaptive stress testing (val), Lipschitz search, DIRECT, simplex, integer programming (opt), decision trees, forests, boosting (learn), backpropagation (diff)
orhof.examples cross-domain compositions that show the patterns interlocking

Full generated reference: docs/API.md. Design notes: ARCHITECTURE.md, docs/SEARCH.md.

The idea

The book's thesis is that these algorithms are not hundreds of unrelated recipes but instances of a few higher-order patterns. Two examples the code makes concrete:

  • One descent framework, many optimizers. Gradient descent, steepest descent, conjugate gradient, and Adam are the same template parameterized by a direction function and a step function.
  • The Bellman operator is a function transformer mdp -> (V -> V'). Value iteration is just its fixpoint; temporal-logic operators are HOFs on specifications that return continuous robustness values.
  • Tree search is one unfold and one fold. Dijkstra, alpha-beta, branch and bound, DIRECT, MCTS, ID3 and backpropagation differ in what they plug in: a successor function, a fold, and for the top-down ones a strategy that orders the frontier. The space is never built.

Testing

clojure -X:test

236 tests. Stochastic algorithms (GA, bandits, PSO, sampling, …) are seeded via a shared fixture, so the suite is deterministic.

Releasing

Push a tag vX.Y.Z. The Release workflow runs the tests and deploys X.Y.Z to Clojars as com.flexiana/orhof. It needs the repository secrets CLOJARS_USERNAME and CLOJARS_PASSWORD (a Clojars deploy token). Locally: clojure -T:build jar or install.

License

MIT — see LICENSE. Copyright © 2026 Jiri Knesl. The companion book is a separate work, licensed CC BY-SA 4.0.

About

Operations research in higher-order functions: a dependency-free Clojure library for optimization, decision making, and validation. Companion to the book 'Applied Higher Order Functions'.

Resources

Stars

20 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages