Heuristic (computer science)

Heuristic (computer science)
Name	Heuristic (computer science)
Caption	Example of heuristic search strategy in a graph
Field	Computer science
Related	Artificial intelligence; Algorithms; Optimization

Heuristic (computer science) Heuristics in computer science are problem-solving methods that employ practical techniques to produce solutions that are good enough within resource constraints, used across Alan Turing-inspired Turing machine research, John von Neumann-style architectures, and modern OpenAI-era systems. They bridge theoretical frameworks from Claude Shannon-led information theory, Edsger Dijkstra-rooted algorithmics, and Donald Knuth-style analysis with applied work at institutions such as Massachusetts Institute of Technology, Stanford University, Carnegie Mellon University, Google, and IBM. Heuristics guide search and decision-making in domains influenced by milestones like the Dartmouth Conference, the AI winter, and the resurgence tied to DeepMind and DARPA programs.

Definition and purpose

Heuristics are practical strategies or rules of thumb developed to find satisfactory solutions when optimal solutions are infeasible due to complexity demonstrated in Cook–Levin theorem and NP-completeness results such as Travelling Salesman Problem instances studied at Bell Labs. They serve purposes in contexts ranging from Bell Labs-style circuit layout problems to NASA mission planning, helping systems like IBM Watson and DeepMind AlphaGo approximate outcomes under constraints set by Moore's law and computational limits at institutions like Intel and NVIDIA. Designed for efficiency, they are evaluated against benchmarks from competitions like those organized by International Joint Conference on Artificial Intelligence and standards promulgated at IEEE conferences.

Types of heuristics

Common types include greedy heuristics used in Dijkstra's algorithm-related contexts, local search heuristics exemplified by Kirkpatrick thermal annealing analogs and Simulated Annealing frameworks developed from thermodynamics work by Metropolis et al., genetic heuristics inspired by Charles Darwin-style selection and popularized by research at Stanford University and University of Michigan, and constraint satisfaction heuristics applied in SAT solving competitions involving groups from MIT and Princeton University. Other classes include admissible heuristics used in A* search implementations at Cornell University and dominance-based heuristics arising from operations research groups at Harvard University and Columbia University.

Design and evaluation

Design draws on theoretical guidance from Richard Karp and empirical methodology used at Bell Labs and AT&T research labs, combining domain knowledge from labs at NASA Jet Propulsion Laboratory and European Space Agency with metrics developed in studies at University of California, Berkeley and University of Oxford. Evaluation uses performance profiles from competitions hosted by ACM and trace-driven experiments modeled after systems at Microsoft Research and Facebook AI Research, comparing solution quality, run time, and resource use against baselines such as exact algorithms demonstrated in work by Garey and Johnson and approximation schemes from Vijay Vazirani.

Applications

Heuristics underpin routing services from Google Maps and logistics platforms at Amazon and FedEx, motion planning in robotics from labs at ETH Zurich and Tokyo Institute of Technology, scheduling in airline operations like those of Delta Air Lines and United Airlines, protein folding approximations pursued by teams at University of Cambridge and Riken, and game-playing engines developed by groups at DeepMind and University of Alberta. They are also critical in compiler optimization phases following principles from Dennis Ritchie and Ken Thompson at Bell Labs, heuristic indexing in databases from Oracle Corporation and SAP, and security threat triage used by NSA and GCHQ.

Limitations and biases

Heuristics can introduce systematic biases exposed in case studies at Stanford and Princeton, analogous to cognitive biases cataloged in work by Daniel Kahneman and Amos Tversky; such biases affected deployments at corporations like Uber and Airbnb and were scrutinized in reviews by European Commission regulators. They may fail catastrophically on worst-case instances identified in theoretical constructs from Cook and Karp, and produce brittle behavior in adversarial contexts explored by researchers at OpenAI and Google DeepMind. Limitations have legal and ethical implications intersecting with policy work at institutions like Harvard Kennedy School and Brookings Institution.

Implementation techniques

Implementation uses data structures and paradigms from seminal work at Bell Labs and MIT CSAIL, leveraging priority queues in A* implementations, tabu lists from metaheuristic research at University of Montreal, population encodings from evolutionary computation groups at EPFL and Georgia Tech, and parallelization strategies deployed on hardware by NVIDIA and Intel. Practical toolchains combine languages and systems influenced by Dennis Ritchie, Bjarne Stroustrup-era C++ compilers, Java runtimes from Sun Microsystems, and Python ecosystems supported by Python Software Foundation and used in projects at OpenAI and Google Brain.

Category:Computer science concepts