"Real-Time Experience-Based Prediction/ Prescription Simulator" (REPS) Technology
REPS is about Time
REPS (Real-time Experience-based Prediction/Prescription Simulator) expands analytics from "the connectedness of things to other things" to "the connectedness of things in time".
Data Preparation
-
Group things "occurring" together at one time (or specific time frame) [note: may not be connected otherwise]
-
Arrange these Groups in time sequence as "Stimulus-Response Pairs" (SRPs)
Machine Learning
-
Train REPS on the SRPs [Note: you may add/remove/modify trained SRPs at any time]
Predictions in Real-Time
-
Given a live situation (i.e. a "stimulus"), REPS uses the trained SRPs to predict behavior of the system (i.e. ordered set of predicted responses). [Note: the stimulus may have never been seen before and/or the response may have never been seen before.]
REPS Prescriptions
REPS applies "Desirability Functions" (DF) to describe the dimensions of relatively better outcomes. For example, a DF could be: minimize cost, minimize casualties, stabilize temperature, etc.
-
The state of the art of Prescriptive Analytics, "...incorporates both structured and unstructured data, and uses a combination of advanced analytic techniques and disciplines to predict, prescribe, and adapt...The technology behind prescriptive analytics synergistically combines hybrid data, business rules with mathematical models and computational models." In general, technical approaches to Prescriptive Analytics technical approaches tend to be cobbled together, ad hoc and widely varied.
-
In contrast, for a given situation REPS locates similar experiences and evaluates their individual and collective outcomes in order to compose changes (ordered change sets) that if applied to the current situation would best maximize the stated order of DFs. Each "change set" can further be mapped to a recommended set of possible actions. Such recommendations can be explained in terms of their contribution to the ordered set of DFs, as well as to previous similar experiences.
REPS Explainable AI
In contrast to statistical approaches to ML, REPS maintains each training experience as a point of reference rather than merge training experience into relative statistics. In REPS, only unique training experiences are retrained, allowing REPS to provide useful results given minimally adequate training data. REPS explanations are in terms of past experiences that support the results as well as the key predicting indicators for aspects of a given context.
-
Explain Prediction: [For each ordered Prediction] (1) List of best supporting Experiences and why; (2) List of Probable Causes and why
-
Explain Prescription: List to what extent each Desirable Outcome Factor is satisfied by each Prescription in the ordered set of Prescriptions; Undesirable Outcome Factors are similarly addressed
Open World vs Closed World
2nd Wave AI is most often "Closed World"; where AI/EBL considers the context and adjusts to an "Open World".
-
Closed: All possible outcomes/solutions are known/knowable
-
Open: New outcomes/solutions are possible
-
Closed: All datatypes are known up front
-
Open: New datatypes can be introduced on the fly
-
Closed: Learn from similar things and similar actions
-
Open: Learn from similar situations (maybe different things and actions)
"Semantic Insights™ Research Assistant" (SIRA) Technology
SIRA for Natural Language Research
SIRA identifies the variety of language patterns and terms in a sentence, and maps them to the concepts and relationships they represent.
Train
-
Starting with a statement of investigation (in natural language), SIRA generates a high-speed "reader program" that can recognize all the ways and terms the investigation's concepts and relationships could be expressed.
Read
-
Then, given a corpus of natural language text, SIRA applies this "reader program" to "read" every sentence, looking for the variety of language patterns and terms that can represent the concepts and relationships expressed in the investigation. All matching information is returned for further processing. SIRA verifies the returned information (linguistically and semantically) for relevance to the investigation.
Report
-
The results can then be rendered as data or natural language in a variety of useful ways.
Harvesting Facts from Natural Language Text
SIRA identiies the same fact expressed using different terms, sentence structure and sentences across multiple natural language texts.
-
Natural Language is a way of encoding facts.
-
Facts, in natural language, are statements about states or relationships of things (e.g. "The red book is heavy." Asserts the "book is red" and the "book on the table is next to the chair,").
-
Facts are assertions we may choose to accept as "true" within a given context.
-
Facts assert the relationship between 2 or more things related in some way.
-
Natural Language provides many different ways to encode the same fact.
-
Regardless of how a fact is stated, when the context is considered, it is still the same fact.
-
So by identifying all the ways a given fact can be expressed in natural language, that fact can be found in any natural language text.
Growing Domain Knowledge from Natural Language Text
OntoGrow is an application within SIRA that specifically exploits meronyms and hypernyms to generate generalization/specialization and composition in the Ontology. In addition, any needed dictionary terms/senses are added.
-
Domain Ontologies: Each domain has a specific ontology (concepts, instances, relationships, generalizations and specializations). When a Language Dictionary is selected, Concepts, Instances and Relationships are usually (but not always) mapped to specific senses in the dictionary (and by implication, the synonyms for that sense).
-
Domain Dictionaries: Each domain has a specific dictionary (set of terms + senses). A sense includes specific linguistic metadata (i.e. how the term is used or behaves), as well as, membership in the set of synonyms senses.