# Inference engine > Mediated Wiki article. Canonical URL: https://mediated.wiki/source/Inference_engine > Markdown URL: https://mediated.wiki/source/Inference_engine.md > Source: https://en.wikipedia.org/wiki/Inference_engine > Source revision: 1347612116 > License: Creative Commons Attribution-ShareAlike 4.0 International (https://creativecommons.org/licenses/by-sa/4.0/) Component of artificial intelligence systems This article needs to be updated. Please help update this article to reflect recent events or newly available information. (October 2019) In the field of [artificial intelligence](/source/Artificial_intelligence), an **inference engine** is a [software component](/source/Software_component) of an intelligent system that applies logical rules to the [knowledge base](/source/Knowledge_base) to deduce new information. The first inference engines were components of [expert systems](/source/Expert_system). The typical expert system consisted of a knowledge base and an inference engine. The knowledge base stored facts about the world. The inference engine applied logical rules to the knowledge base and deduced new knowledge. This process would iterate as each new fact in the knowledge base could trigger additional rules in the inference engine. Inference engines work primarily in one of two modes either special rule or facts: [forward chaining](/source/Forward_chaining) and [backward chaining](/source/Backward_chaining). Forward chaining starts with the known facts and asserts new facts. Backward chaining starts with goals, and works backward to determine what facts must be asserted so that the goals can be achieved.[1] Additionally, the concept of 'inference' has expanded to include the process through which trained [neural networks](/source/Artificial_neural_network) generate predictions or decisions. In this context, an 'inference engine' could refer to the specific part of the system, or even the hardware, that executes these operations. This type of inference plays a crucial role in various applications, including (but not limited to) [image recognition](/source/Image_recognition), [natural language processing](/source/Natural_language_processing), and [autonomous vehicles](/source/Autonomous_vehicles). The inference phase in these applications is typically characterized by a high volume of data inputs and real-time processing requirements. ## Architecture The logic that an inference engine uses is typically represented as IF-THEN rules. The general format of such rules is IF THEN . Prior to the development of expert systems and inference engines, artificial intelligence researchers focused on more powerful [theorem prover](/source/Theorem-prover) environments that offered much fuller implementations of [first-order logic](/source/First-order_logic). For example, general statements that included [universal quantification](/source/Universal_quantification) (for all X some statement is true) and [existential quantification](/source/Existential_quantification) (there exists some X such that some statement is true). What researchers discovered is that the power of these theorem-proving environments was also their drawback. Back in 1965, it was far too easy to create logical expressions that could take an indeterminate or even infinite time to terminate. For example, it is common in universal quantification to make statements over an infinite set such as the set of all natural numbers. Such statements are perfectly reasonable and even required in mathematical proofs but when included in an automated theorem prover executing on a computer may cause the computer to fall into an infinite loop. Focusing on IF-THEN statements (what logicians call *[modus ponens](/source/Modus_ponens)*) still gave developers a very powerful general mechanism to represent logic, but one that could be used efficiently with computational resources. What is more, there is some psychological research that indicates humans also tend to favor IF-THEN representations when storing complex knowledge.[2] A simple example of *modus ponens* often used in introductory logic books is "If you are human then you are mortal". This can be represented in [pseudocode](/source/Pseudocode) as: Rule1: Human(x) => Mortal(x) A trivial example of how this rule would be used in an inference engine is as follows. In *forward chaining*, the inference engine would find any facts in the knowledge base that matched Human(x) and for each fact it found would add the new information Mortal(x) to the knowledge base. So if it found an object called Socrates that was human it would deduce that Socrates was mortal. In *backward chaining*, the system would be given a goal, e.g. answer the question is Socrates mortal? It would search through the knowledge base and determine if Socrates was human and, if so, would assert he is also mortal. However, in backward chaining a common technique was to integrate the inference engine with a user interface. In that way, rather than simply being automated the system could now be interactive. In this trivial example, if the system was given the goal to answer the question if Socrates was mortal and it didn't yet know if he was human, it would generate a window to ask the user the question "Is Socrates human?" and would then use that information accordingly. This innovation of integrating the inference engine with a user interface led to the second early advancement of expert systems: explanation capabilities. The explicit representation of knowledge as rules rather than code made it possible to generate explanations to users: both explanations in real time and after the fact. So if the system asked the user "Is Socrates human?", the user may wonder why she was being asked that question and the system would use the chain of rules to explain why it was currently trying to ascertain that bit of knowledge: that is, it needs to determine if Socrates is mortal and to do that needs to determine if he is human. At first these explanations were not much different than the standard debugging information that developers deal with when debugging any system. However, an active area of research was utilizing natural language technology to ask, understand, and generate questions and explanations using natural languages rather than computer formalisms.[3] An inference engine cycles through three sequential steps: *match rules*, *select rules*, and *execute rules*. The execution of the rules will often result in new facts or goals being added to the knowledge base, which will trigger the cycle to repeat. This cycle continues until no new rules can be matched. In the first step, *match rules*, the inference engine finds all of the rules that are triggered by the current contents of the knowledge base. In forward chaining, the engine looks for rules where the antecedent (left hand side) matches some fact in the knowledge base. In backward chaining, the engine looks for antecedents that can satisfy one of the current goals. In the second step, *select rules*, the inference engine prioritizes the various rules that were matched to determine the order to execute them. In the final step, *execute rules*, the engine executes each matched rule in the order determined in step two and then iterates back to step one again. The cycle continues until no new rules are matched.[4] ## Implementations Early inference engines focused primarily on forward chaining. These systems were usually implemented in the [Lisp](/source/Lisp_(programming_language)) programming language. Lisp was a frequent platform for early AI research due to its strong capability to do symbolic manipulation. Also, as an [interpreted language](/source/Interpreted_language) it offered productive development environments appropriate to [debugging](/source/Debugging) complex programs. A necessary consequence of these benefits was that Lisp programs tended to be slower and less robust than compiled languages of the time such as [C](/source/C_(programming_language)). A common approach in these early days was to take an expert system application and repackage the inference engine used for that system as a re-usable tool other researchers could use for the development of other expert systems. For example, [MYCIN](/source/MYCIN) was an early expert system for medical diagnosis and EMYCIN was an inference engine extrapolated from MYCIN and made available for other researchers.[1] As expert systems moved from research prototypes to deployed systems there was more focus on issues such as speed and robustness. One of the first and most popular forward chaining engines was [OPS5](/source/OPS5), which used the [Rete algorithm](/source/Rete_algorithm) to optimize the efficiency of rule firing. Another very popular technology that was developed was the [Prolog](/source/Prolog) logic programming language. Prolog focused primarily on backward chaining and also featured various commercial versions and optimizations for efficiency and robustness.[5] As expert systems prompted significant interest from the business world, various companies, many of them started or guided by prominent AI researchers created productized versions of inference engines. For example, [Intellicorp](/source/IntelliCorp_(software)) was initially guided by [Edward Feigenbaum](/source/Edward_Feigenbaum). These inference engine products were also often developed in Lisp at first. However, demands for more affordable and commercially viable platforms eventually made [personal computer](/source/Personal_computer) platforms very popular. ### Open source implementations [ClipsRules](https://clipsrules.net/) and [RefPerSys](http://refpersys.org) (inspired by [CAIA](https://github.com/bstarynk/caia-pitrat)[6] and the work of [Jacques Pitrat](/source/Jacques_Pitrat)). The [Frama-C](https://frama-c.com/) static source code analyzer also uses some inference engine techniques. ## See also - [Geometric and Topological Inference](/source/Geometric_and_Topological_Inference) - [Action selection](/source/Action_selection) - [Backward chaining](/source/Backward_chaining) - [Expert system](/source/Expert_system) - [Forward chaining](/source/Forward_chaining) - [Inductive inference](/source/Inductive_inference) ## References 1. ^ [***a***](#cite_ref-Hayes-Roth_1983_1-0) [***b***](#cite_ref-Hayes-Roth_1983_1-1) Hayes-Roth, Frederick; Donald Waterman; Douglas Lenat (1983). [*Building Expert Systems*](https://archive.org/details/buildingexpertsy00temd). Addison-Wesley. [ISBN](/source/ISBN_(identifier)) [0-201-10686-8](https://en.wikipedia.org/wiki/Special:BookSources/0-201-10686-8).[*[page needed](https://en.wikipedia.org/wiki/Wikipedia:Citing_sources)*] 1. **[^](#cite_ref-2)** Feigenbaum, Edward; Avron Barr (September 1, 1986). [*The Handbook of Artificial Intelligence, Volume I*](https://archive.org/stream/handbookofartific01barr#page/156/mode/2up). Addison-Wesley. p. 195. [ISBN](/source/ISBN_(identifier)) [0201118114](https://en.wikipedia.org/wiki/Special:BookSources/0201118114). 1. **[^](#cite_ref-3)** Barzilayt, Regina; McCullough, Daryl; Rambow, Owen; DeCristofaro, Jonathan; Korelsky, Tanya; Lavoie, Benoit. "A New Approach to Expert System Explanations". *USAF Rome Laboratory Report*. [DTIC](/source/DTIC_(identifier)) [ADA457707](https://apps.dtic.mil/sti/citations/tr/ADA457707). 1. **[^](#cite_ref-4)** Griffin, N.L.; Lewis, F.D. (1989). "A rule-based inference engine which is optimal and VLSI implementable". *[Proceedings 1989] IEEE International Workshop on Tools for Artificial Intelligence*. pp. 246–251. [Bibcode](/source/Bibcode_(identifier)):[1989tai..conf...32G](https://ui.adsabs.harvard.edu/abs/1989tai..conf...32G). [doi](/source/Doi_(identifier)):[10.1109/TAI.1989.65327](https://doi.org/10.1109%2FTAI.1989.65327). [ISBN](/source/ISBN_(identifier)) [0-8186-1984-8](https://en.wikipedia.org/wiki/Special:BookSources/0-8186-1984-8). 1. **[^](#cite_ref-5)** Sterling, Leon; Ehud Shapiro (1986). [*The Art of Prolog*](https://archive.org/details/artofprologadvan00ster). Cambridge, MA: MIT. [ISBN](/source/ISBN_(identifier)) [0-262-19250-0](https://en.wikipedia.org/wiki/Special:BookSources/0-262-19250-0).[*[page needed](https://en.wikipedia.org/wiki/Wikipedia:Citing_sources)*] 1. **[^](#cite_ref-6)** Pitrat, Jacques (2009). *Artificial Beings, the conscience of a conscious machine*. Wiley. [ISBN](/source/ISBN_(identifier)) [978-1848211018](https://en.wikipedia.org/wiki/Special:BookSources/978-1848211018).[*[page needed](https://en.wikipedia.org/wiki/Wikipedia:Citing_sources)*] Authority control databases GND v t e Knowledge representation and reasoning Knowledge bases Cyc ConceptNet DBpedia Knowledge Graph YAGO Wikidata Reasoning systems Expert systems CLIPS Connectionist expert systems Expert systems for mortgages Legal expert systems Theorem provers CARINE E Nqthm Otter Paradox Prover9 SPASS TPS Z3 Logic programs Prolog Datalog ASP Constraint solvers Constraint Handling Rules MiniZinc OR-Tools Others Deductive classifiers Procedural reasoning systems Rule engines Ontology languages Attempto Controlled English CycL F-logic FO(.) Knowledge Interchange Format Web Ontology Language Constraint satisfaction Constraint programming Constraint logic programming Local consistency SMT solvers Automated planning Motion planning Multi-agent planning Partial-order planning Preference-based planning Reactive planning State-space planning Argumentation frameworks Backward chaining Blackboard systems Case-based reasoning Commonsense reasoning Forward chaining Model-based reasoning Inference engines Knowledge engineering Knowledge acquisition Knowledge extraction Rule-based systems Spatio-temporal reasoning Qualitative reasoning --- Adapted from the Wikipedia article [Inference engine](https://en.wikipedia.org/wiki/Inference_engine) by Wikipedia contributors ([contributor history](https://en.wikipedia.org/wiki/Inference_engine?action=history)). Available under [Creative Commons Attribution-ShareAlike 4.0 International](https://creativecommons.org/licenses/by-sa/4.0/). Changes may have been made.