Hybrid Machine Learning / Personality Simulation Platform
"Towards an Event-Based Simulation/ML Hybrid Platform"
Abstract: Machine learning systems have become the dominant form of AI for enterprises due in large part to years of experimentation with combinations of technical approaches. Today’s ML systems are unlike traditional software systems in that algorithms are not used to procedurally specify how systems learn to recognize patterns from input. On the other hand, simulation systems have traditionally been highly procedural, which is good for understanding how a result was obtained, but ‘learning’ is not normally expected. Mike Slinn will discuss a hybrid approach under consideration based on work he started in 2008.
Bio: Companies that Mike Slinn cofounded, led and advised have been sold to IBM, Otsuka, Microsoft, Yahoo! and NBC Interactive. A recognized software expert in US and European courts, Mike opines on contractual and patent disputes. Mike received an Electronics Engineering degree in 1979 from Carleton University in Ottawa.
HOW IT WORKS
Because EmpathyWorks uses event sourcing:
An individual’s emotional state can be determined at any point in time.
The factors that went into a prediction can be identified and manipulated.
By considering multiple timelines, what-if scenarios can be be forecast.
The "best" algorithm or type of processing is contextually dependent. EmpathyWorks will feature pluggable computation modes in a configurable data flow. Extremely high volumes of data will be processed.
EmpathyWorks is real-time event-driven, and events for many or most installations are likely to be provided by machine learning systems that use classifications and regressions. Internal computations within the data flow might be defined by algorithms or employ machine learning.
The Future is Now
Mike Slinn started thinking deeply about artificial personality as a an undergraduate engineering student. Decades later, in 2008, he wrote the first EmpathyWorks prototype, a configurable behavioral modeler for one organism. Several years later the second prototype focused on group dynamics. Now it is time to build this technology into a general-purpose platform while considering the markets that such a product might address. Would you like to get involved?
We are moving from many years of R&D, and as of January 2020 we will begin a 6-month market research phase, which will be followed by a consultative phase, then a productization phase. EmpathyWorks began as a single-person project; successive phases will transition the project to an ever-larger company. The end state will have us functioning as a tech vendor / partner.
Simulating behavior would be useful for breathing life into video game characters, computer generated characters in movies, virtual pets and artificial friends.
Predicting behavior of individuals and groups would lead to better understanding the factors that led to observed outcomes. Understanding the tipping points for various outcomes would allow for better risk management for sensitive situations - a more graceful and harmonious world.
Originally written in Java with .NET runtime support, the most recent code bases is written in Scala. Some experimentation has been done with Apache Kafka and CRDTs. Various options for integrating with machine learning platforms such as Apache Spark need to be explored, including embedding and supporting plugin compatibility.
While Adobe Flex was orignally used for the user interface, that technlogy is now dead. Some of the current user interface considerations are incredibly exciting, for example rendering deep fakes.
Declarative – the Personality Rule Language (PRL™) defines the model declaratively
Live reloading – the EmpathyWorks meta-model can be updated while the system is running without loss of data.
Flexible – the model can define arbitrary species and individuals, behavioral characteristics, environment and relationships
Device Independent – a companion AI application informs EmpathyWorks of events and EmpathyWorks simulates the response of the personalities
Time sensitive – simulated personalities respond to the passage of time
Scalable – Could model the entire world‘s population
Multi-generational inheritance – personality traits can be inherited from parents
Event types – Responses to life-changing events are defined separately from mundane events and environmental events
Societal modeling – relationships between individuals can be defined
Shared events – one artificial personality’s life events can affect the other personalities with which it has relationships with
Predator/prey relationships – behaviors corresponding to hunting and being hunted are supported
Decision making – an application can query an artificial personality for its opinions as to which of several choices to make
Life stages – artificial personalities can grow up to mature individuals, mature, become old and die
Emotional information – an artificial personality’s internal state can be queried for display