Solvor all your optimization needs.

What's in the box?

Category	Solvors	Use Case
Linear/Integer	solve_lp, solve_milp	Resource allocation, scheduling
Constraint	solve_sat, Model	Sudoku, puzzles, and that one config problem that's been bugging you
Combinatorial	solve_knapsack, solve_bin_pack, solve_job_shop, solve_vrptw	Packing, scheduling, routing
Local Search	anneal, tabu_search, lns, alns	TSP, combinatorial optimization
Population	evolve, differential_evolution, particle_swarm	Global search, nature-inspired
Gradient	gradient_descent, momentum, rmsprop, adam	ML, curve fitting
Quasi-Newton	bfgs, lbfgs	Fast convergence, smooth functions
Derivative-Free	nelder_mead, powell, bayesian_opt	Black-box, expensive functions
Pathfinding	bfs, dfs, dijkstra, astar, astar_grid, bellman_ford, floyd_warshall	Shortest paths, graph traversal
Graph	max_flow, min_cost_flow, kruskal, prim	Flow, MST, connectivity
Assignment	solve_assignment, solve_hungarian, network_simplex	Matching, min-cost flow
Exact Cover	solve_exact_cover	N-Queens, tiling puzzles
Utilities	FenwickTree, UnionFind	Data structures for algorithms

---

Quickstart

uv add solvor

from solvor import solve_lp, solve_tsp, dijkstra, solve_hungarian

# Linear Programming
result = solve_lp(c=[1, 2], A=[[1, 1], [2, 1]], b=[4, 5])
print(result.solution)  # optimal x

# TSP with tabu search
distances = [[0, 10, 15], [10, 0, 20], [15, 20, 0]]
result = solve_tsp(distances)
print(result.solution)  # best tour found

# Shortest path
graph = {'A': [('B', 1), ('C', 4)], 'B': [('C', 2)], 'C': []}
result = dijkstra('A', 'C', lambda n: graph.get(n, []))
print(result.solution)  # ['A', 'B', 'C']

# Assignment
costs = [[10, 5], [3, 9]]
result = solve_hungarian(costs)
print(result.solution)  # [1, 0] - row 0 gets col 1, row 1 gets col 0

---

Solvors

Linear & Integer Programming

solve_lp

For resource allocation, blending, production planning. Finds the exact optimum for linear objectives with linear constraints.

# minimize 2x + 3y subject to x + y >= 4, x <= 3
result = solve_lp(
    c=[2, 3],
    A=[[-1, -1], [1, 0]],  # constraints as Ax <= b
    b=[-4, 3]
)

solve_milp

When some variables must be integers. Diet problems, scheduling with discrete slots, set covering.

# same as above, but x must be integer
result = solve_milp(c=[2, 3], A=[[-1, -1], [1, 0]], b=[-4, 3], integers=[0])

Constraint Programming

solve_sat

For "is this configuration valid?" problems. Dependencies, exclusions, implications - anything that boils down to boolean constraints.

# (x1 OR x2) AND (NOT x1 OR x3) AND (NOT x2 OR NOT x3)
result = solve_sat([[1, 2], [-1, 3], [-2, -3]])
print(result.solution)  # {1: True, 2: False, 3: True}

Model

Constraint programming for puzzles and scheduling with "all different", arithmetic, and logical constraints. Sudoku, N-Queens, timetabling.

m = Model()
cells = [[m.int_var(1, 9, f'c{i}{j}') for j in range(9)] for i in range(9)]

# All different in each row
for row in cells:
    m.add(m.all_different(row))

result = m.solve()

Metaheuristics

anneal

Simulated annealing, sometimes you have to go downhill to find a higher peak.

result = anneal(
    initial=initial_solution,
    objective_fn=cost_function,
    neighbors=random_neighbor,
    temperature=1000,
    cooling=0.9995
)

tabu_search

Greedy local search with a "don't go back there" list. Simple but surprisingly effective.

result = tabu_search(
    initial=initial_solution,
    objective_fn=cost_function,
    neighbors=get_neighbors,  # returns [(move, solution), ...]
    cooldown=10
)

lns / alns

Large Neighborhood Search, destroy part of your solution and rebuild it better. ALNS learns which operators work best.

result = lns(initial, objective_fn, destroy_ops, repair_ops, max_iter=1000)
result = alns(initial, objective_fn, destroy_ops, repair_ops, max_iter=1000)  # adaptive

evolve / differential_evolution / particle_swarm

Population-based global search. Let a swarm of candidates explore for you. DE and PSO work on continuous spaces.

result = evolve(objective_fn=fitness, population=pop, crossover=cx, mutate=mut)
result = differential_evolution(objective_fn, bounds=[(0, 10)] * n, population_size=50)
result = particle_swarm(objective_fn, bounds=[(0, 10)] * n, n_particles=30)

Continuous Optimization

Gradient-Based

gradient_descent, momentum, rmsprop, adam - follow the slope. Adam adapts learning rates per parameter.

def grad_fn(x):
    return [2 * x[0], 2 * x[1]]  # gradient of x^2 + y^2

result = adam(grad_fn, x0=[5.0, 5.0])

Quasi-Newton

bfgs, lbfgs - approximate Hessian for faster convergence. L-BFGS uses limited memory.

result = bfgs(objective_fn, grad_fn, x0=[5.0, 5.0])
result = lbfgs(objective_fn, grad_fn, x0=[5.0, 5.0], memory=10)

Derivative-Free

nelder_mead, powell - no gradients needed. Good for noisy, non-smooth, or "I have no idea what this function looks like" situations.

result = nelder_mead(objective_fn, x0=[5.0, 5.0])
result = powell(objective_fn, x0=[5.0, 5.0])

bayesian_opt

When each evaluation is expensive. Builds a surrogate model to sample intelligently. Perfect for hyperparameter tuning or when your simulation takes 10 minutes per run.

result = bayesian_opt(expensive_fn, bounds=[(0, 1), (0, 1)], max_iter=30)

Pathfinding

bfs / dfs

Unweighted graph traversal. BFS finds shortest paths (fewest edges), DFS explores depth-first. Both work with any state type and neighbor function.

# Find shortest path in a grid
def neighbors(pos):
    x, y = pos
    return [(x+1, y), (x-1, y), (x, y+1), (x, y-1)]

result = bfs(start=(0, 0), goal=(5, 5), neighbors=neighbors)
print(result.solution)  # path from start to goal

dijkstra

Weighted shortest paths. Classic algorithm for when edges have costs. Stops early when goal is found.

def neighbors(node):
    return graph[node]  # returns [(neighbor, cost), ...]

result = dijkstra(start='A', goal='Z', neighbors=neighbors)

astar / astar_grid

A* with heuristics. Faster than Dijkstra when you have a good distance estimate. astar_grid handles 2D grids with built-in heuristics.

# Grid pathfinding with obstacles
grid = [[0, 0, 1, 0], [0, 0, 1, 0], [0, 0, 0, 0]]
result = astar_grid(grid, start=(0, 0), goal=(2, 3))

bellman_ford

Handles negative edge weights. Slower than Dijkstra but detects negative cycles, which is useful when costs can go negative.

result = bellman_ford(start=0, edges=[(0, 1, 5), (1, 2, -3), ...], n_nodes=4)

floyd_warshall

All-pairs shortest paths. O(n³) but gives you every shortest path at once. Worth it when you need the full picture.

result = floyd_warshall(n_nodes=4, edges=[(0, 1, 3), (1, 2, 1), ...])
# result.solution[i][j] = shortest distance from i to j

Network Flow & MST

max_flow

"How much can I push through this network?" Assigning workers to tasks, finding bottlenecks.

graph = {
    's': [('a', 10, 0), ('b', 5, 0)],
    'a': [('b', 15, 0), ('t', 10, 0)],
    'b': [('t', 10, 0)],
    't': []
}
result = max_flow(graph, 's', 't')

min_cost_flow / network_simplex

"What's the cheapest way to route X units?" Use min_cost_flow for simplicity, network_simplex for large instances.

# network_simplex for large min-cost flow
arcs = [(0, 1, 10, 2), (0, 2, 5, 3), (1, 2, 15, 1)]  # (from, to, cap, cost)
supplies = [10, 0, -10]  # positive = source, negative = sink
result = network_simplex(n_nodes=3, arcs=arcs, supplies=supplies)

kruskal / prim

Minimum spanning tree. Connect all nodes with minimum total edge weight. Kruskal sorts edges, Prim grows from a start node.

edges = [(0, 1, 4), (0, 2, 3), (1, 2, 2)]  # (u, v, weight)
result = kruskal(n_nodes=3, edges=edges)
# result.solution = [(1, 2, 2), (0, 2, 3)] - MST edges

Assignment

solve_assignment / solve_hungarian

Optimal one-to-one matching. solve_hungarian is O(n³), direct algorithm for assignment problems.

costs = [
    [10, 5, 13],
    [3, 9, 18],
    [10, 6, 12]
]
result = solve_hungarian(costs)
# result.solution[i] = column assigned to row i
# result.objective = total cost

# For maximization
result = solve_hungarian(costs, minimize=False)

Exact Cover

solve_exact_cover

For "place these pieces without overlap" problems. Sudoku, pentomino tiling, N-Queens, or any puzzle where you stare at a grid wondering why nothing fits.

# Tiling problem: cover all columns with non-overlapping rows
matrix = [
    [1, 1, 0, 0],  # row 0 covers columns 0, 1
    [0, 1, 1, 0],  # row 1 covers columns 1, 2
    [0, 0, 1, 1],  # row 2 covers columns 2, 3
    [1, 0, 0, 1],  # row 3 covers columns 0, 3
]
result = solve_exact_cover(matrix)
# result.solution = (0, 2) or (1, 3) - rows that cover all columns exactly once

Combinatorial Solvers

solve_knapsack

The classic "what fits in your bag" problem. Select items to maximize value within capacity.

values = [60, 100, 120]
weights = [10, 20, 30]
result = solve_knapsack(values, weights, capacity=50)
# result.solution = [1, 1, 1] - which items to take

solve_bin_pack

Bin packing heuristics. Minimize bins needed for items.

items = [4, 8, 1, 4, 2, 1]
result = solve_bin_pack(items, bin_capacity=10)
# result.solution = (1, 0, 0, 1, 0, 0) - bin index for each item
# result.objective = 2 (number of bins)

solve_job_shop

Job shop scheduling. Minimize makespan for jobs on machines.

# jobs[i] = [(machine, duration), ...] - operations for job i
jobs = [[(0, 3), (1, 2)], [(1, 2), (0, 4)]]
result = solve_job_shop(jobs, n_machines=2)

solve_vrptw

Vehicle Routing with Time Windows. Serve customers with capacity and time constraints.

from solvor import Customer, solve_vrptw

customers = [Customer(x=0, y=0, demand=0, ready=0, due=100, service=0)]  # depot
customers += [Customer(x=i, y=i, demand=10, ready=0, due=50, service=5) for i in range(1, 5)]
result = solve_vrptw(customers, vehicle_capacity=30, n_vehicles=2)

---

Result Format

All solvors return a consistent Result dataclass:

Result(
    solution,     # best solution found
    objective,    # objective value
    iterations,   # solver iterations (pivots, generations, etc.)
    evaluations,  # function evaluations
    status,       # OPTIMAL, FEASIBLE, INFEASIBLE, UNBOUNDED, MAX_ITER
    error,        # error message if failed (None on success)
    solutions,    # multiple solutions when solution_limit > 1
)

---

When to use what?

Problem	Solvor
Linear constraints, continuous	solve_lp
Linear constraints, integers	solve_milp
Boolean satisfiability	solve_sat
Discrete vars, complex constraints	Model
Knapsack, subset selection	solve_knapsack
Bin packing	solve_bin_pack
Job shop scheduling	solve_job_shop
Vehicle routing	solve_vrptw
Combinatorial, local search	tabu_search, anneal
Combinatorial, destroy-repair	lns, alns
Global search, continuous	differential_evolution, particle_swarm
Global search, discrete	evolve
Smooth, differentiable, fast	bfgs, lbfgs
Smooth, differentiable, ML	adam, rmsprop
Non-smooth, no gradients	nelder_mead, powell
Expensive black-box	bayesian_opt
Shortest path, unweighted	bfs, dfs
Shortest path, weighted	dijkstra, astar
Shortest path, negative weights	bellman_ford
All-pairs shortest paths	floyd_warshall
Minimum spanning tree	kruskal, prim
Maximum flow	max_flow
Min-cost flow	min_cost_flow, network_simplex
Assignment, matching	solve_hungarian, solve_assignment
Exact cover, tiling	solve_exact_cover

---

Philosophy

1. Pure Python: no numpy, no scipy, no compiled extensions. Copy it, change it, break it, learn from it.
2. Readable: each solvor fits in one file you can actually read
3. Consistent: same Result format, same minimize/maximize convention
4. Practical: solves real problems (and AoC puzzles, obviously)

---

Documentation

Full docs at solvOR.ai: getting started, algorithm reference, cookbook with worked examples, and troubleshooting.

---

Contributing

See CONTRIBUTING.md for development setup and guidelines.

License

Apache 2.0 License, free for personal and commercial use.

Background of solvOR

A little bit of history..

I learned about solvers back in 2011, working with some great minds. I started writing python myself around 2018, always as a hobby, and in 2024 I got back into solvers for an energy management system (EMS) and wrote a few tools (simplex, milp, genetic) myself mainly to improve my understanding.

Over time this toolbox got larger and larger, so I decided to publish it on GitHub so others can use it and improve it even further. Since I am learning Rust, I will eventually replace some performance critical operations with a high performance Rust implementation. But since I work on this project (and others) in my spare time, what and when is uncertain. The name solvOR is a mix of solver(s) and OR (Operations Research).

Disclaimer; I am not a professional software engineer, so there are probably some obvious improvements possible. If so let me know, or create a PR!