Encode directed coincidences
For a permutation, x→y is an adjacent ordered pair. The representation is deliberately local and position-independent.
Thesis Revisited · Research archive in development
COIN learns from what works—and what repeatedly fails.
The Coincidence Algorithm is a probabilistic approach to combinatorial optimization. It studies useful partial structures in strong solutions while reducing structures repeatedly observed in poor ones.
Editorial status: This 2026 collection is reconstructed from the completed 2010 doctoral dissertation, Hybrid Positive and Negative Correlation Learning in Estimation of Distribution Algorithm for Combinatorial Optimization Problems. Historical findings and present-day interpretation are labelled separately.
Evidence entering the model
COIN / 01
+
Useful structuresIncrease the probability of patterns observed in stronger candidates.
probability model
−
Repeated poor structuresReduce the probability of patterns associated with weaker candidates.
sample → evaluate → learn → sample again
The research question
Can an optimizer learn from rejected solutions without treating every failure as the same kind of evidence?
Optimization is usually described as a search for better answers. I see it as a disciplined process of deciding which structures deserve to survive, which structures deserve suspicion, and how much certainty the available evidence can support.
COIN begins with a familiar premise: good candidate solutions often share useful partial structures. It then asks a less comfortable question. If certain relationships recur in poor candidates, should the model continue sampling them as if nothing had been learned?
Architecture at a glance
A population proposes evidence. H remembers it.
COIN replaces explicit crossover with a first-order probability model. Each row of H describes what may follow one state; better and worse candidates move probability in opposite directions while every row remains normalised.
Turn candidates into evidence
The objective function ranks the sampled population; multi-objective problems use dominance and diversity information.
Reward and punish
Edges recurring in stronger candidates gain probability. Edges recurring in weaker candidates surrender probability to alternatives.
Construct feasible permutations
The sampler walks through H, masks already-used states, renormalises feasible successors and repeats the cycle.
Research orientation
One algorithm, three ways to enter the work
The collection is being rebuilt for readers who need different levels of detail without weakening the underlying research record.
Understand the idea
Begin with an intuitive account of positive and negative evidence, partial order and probability-guided search.
Read the research motivation →Inspect the mechanism
Follow the representation, conditional model, reward and punishment updates, sampling cycle and multi-objective extension.
Inspect the architecture →Question the evidence
Review benchmark results, comparisons, limitations and the industrial worker-allocation case without hiding uncertainty.
Review the experiments →2026 collection architecture
The COIN research map
The updated dissertation is now the source of record. Each destination translates one part of its argument into a contemporary teaching and research narrative while retaining its original scope.
01
Optimization FoundationsPermutation spaces, neighborhoods, recombination, metaheuristics and probabilistic model-building.
Chapters I–II
02
Negative KnowledgeWhat failure can teach a model—and when that signal becomes unreliable.
Research concept
03
Algorithm ArchitectureRepresentation, model updates, sampling, computational cost and extensions.
Technical
03B
Multi-Objective & Multimodal COINHow one shared generator prunes common negative building blocks while retaining the useful structures of diverse solution modes.
Capability dossier
04
Empirical AnalysisMagic squares, chess-placement puzzles and knight’s tours used to isolate how negative learning behaves.
Evidence
05
Real-World ApplicationsTravelling salesperson benchmarks, U-shaped line balancing and production sequencing.
Industry
06
Metaheuristic Learning PathwayLocal search, genetic algorithms, ACO, PSO, EDAs and multi-objective search.
Learning pathway
Open research software · 2026
From dissertation record to reusable COIN laboratories.
The current Python suite separates reusable algorithms from applications, preserves scalar and Pareto selection, and makes Edge, NB, CNB, Hybrid, EHBSA and NHBSA variants available under shared permutation-problem contracts.
COINCIDENCE Algorithms Suite
Library-only Python source, tests, citation metadata, TSP/MO-TSP notebooks and reusable evaluators.
Open the GitHub repository →Permutation research laboratories
Run Knight's Tour, Flow Shop, TSP with Time Windows, Sudoku and RNA experiments with explicit evaluation budgets.
Explore live COIN laboratories →Multi-objective COIN
Inspect nondominated archives, objective-wise learning progress, convergence, spread and trade-offs across algorithms.
Study the MO research map →Recent COIN work · live laboratories
One reusable library, several combinatorial problems.
These pages connect the historical research record to current executable work. Each laboratory exposes representation, parameters, objective-evaluation budget and learning progress rather than presenting only a final answer.
KTKnight's TourVerified 63-move semantics, learned edge coincidence and open/closed-tour exploration.Numba
FSPermutation Flow ShopTaillard scheduling, scalar and Pareto objectives, COIN/EDA/evolutionary comparisons.Scheduling
TWTSP with Time WindowsTravel cost, makespan, waiting and lateness under hard or soft windows.Routing
SUSudoku as OptimizationBlock-permutation encoding, prime-product and decomposed multi-objective guidance.Constraint search
RNARNA Secondary StructureCompatible helix selection, proxy/exact evaluation and inspectable structure views.Bioinformatics
PYCOINCIDENCE Algorithms SuiteLibrary source, tests, notebooks, citation metadata and reusable problem contracts.GitHub
Research record
Source status is part of the evidence.
The source of record is the 172-page doctoral dissertation completed in the Department of Computer Engineering, Chulalongkorn University, in academic year 2010 under the supervision of Professor Prabhas Chongstitvatana. The web edition interprets that research; it does not replace the dissertation.