What is Information? What is Intelligence?
Building Artificial General Intelligence (AGI) Agents
We developed the Social AI Platform and the Information Digital Twin (IDT) to provide commercial AGI capabilities and solutions. As such, the Social AI Platform is intended to support a wide range of intelligent scenarios using the same agent: the Information Digital Twin (IDT).
The Social AI Platform technology is the realization of our novel concepts and approaches to information and intelligence.
In its mature form, the IDT can be considered as a general reinforcement learning (RL) agent
Information and Intelligence-Unified Definitions*
We believe that the path towards building Artificial General Intelligence (AGI) agents is based on unified-and dependent-definitions of information and intelligence
Artificial approaches to intelligence depend on computational models to process information and provide intelligent capabilities. Due to the lack of uniform definitions of what constitutes intelligence and what is information, the capabilities such models provide differ according to their interpretations of intelligence
and implicit assumptions about what is information. The variety of interpretations of intelligence and information also indicate that existing intelligence computational models provide specialized rather than general capabilities.
The following are the unified definitions we have developed so far and use to mature the IDT and the Social AI Platform.
A system's degree of entanglement with its environment measured in bits. That is, its capacity to predict its environment and be predictable by it
Information (and intelligence) are apparent through interactions. If an agent or a system does not interact with its environment, it is impossible to tell if it processes information or if it is intelligent. Interaction, or more precisely communication, is thus a prerequisite for understanding and defining concepts like information and intelligence.
To communicate, an agent relies on a set of symbols, alphabet, or code and some rules to construct messages and to send them to its environment, and to deconstruct the messages it receives from other agents.
If the various agents depend on one another in reaching their goals, then the set of symbols used by each agent to build its messages will start developing some degree of correlations. Although each agent might use a completely arbitrary set of symbols, they need to make sure to use these symbols consistently in order to convey their messages effectively. This consistency is what achieves the correlations among the various sets of symbols.
Once correlations are established, predictability follows. When an agent sends a specific message, it predicts receiving a particular return message. In the other direction, when an agent receives a message, the level of correspondence of its response message determines the environment's capacity to predict its behavior.
Our research assumption is that the level of correlations and the resulting agent-environment predictability is a unique, general property of a system interacting with its environment, similar to properties such as energy or mass. Consequently, we attempt to demonstrate that information is the metric for measuring this level of correlations, or entanglement. To formalize the information metric, we then rely on concepts from communication theory.
The activities of a system to maximize its entanglement with its environment, That is, the system strive to maximize its information, and accordingly its capacity to predict and be predictable
The relationship between intelligence and information, or level of predictability as we define it here, is rooted in the nature of communication.
Communication theory (Shannon and Weaver, 1949) defines communication as "the procedures by which one mind may affect another." To include "non-mental" systems, communication is thus the procedure by which one entity influences the actions of another in a particular direction. An example is a thermostat sending a control signal to a heater to change its degree of transferring energy into heat.
Developing an effective communication system is a highly complex task. An agent needs to define a limited set of unique symbols, define rules to structure them into signals and messages (syntax), and map them to the set of concepts necessary to achieve its objectives (semantics).
However, the most challenging aspect of communication is the activity of maintaining, updating, and optimizing the symbols, syntax, and semantics rules to achieve effective communication. This aspect is where our research assumption about the nature of intelligence comes into play.
Our research assumption is that intelligence is the continuous activity of an agent to evaluate its communication effectiveness and increase its capacity to influence its environment toward achieving its objectives. This translates into monitoring the usage of its communication symbols, updating the rules to build messages out of these symbols, updating the messages to reflect its objectives, and ultimately evaluating its overall success in influencing its environment or adapting to it.
The question here is, what is the control signal that helps the system to accomplish these activities? Our assumption is that one possible way for the system to measure its effectiveness in completing these tasks is its overall information: its ability to predict and be predictable. That is, the intelligence of an agent can be assessed by its predictability capacity and if it can maintain it or increase it against changes in its environment.
Research Opportunities: Integrating & Managing Heterogeneous Systems
One method to substantiate the definition of information as a measure of entanglement is to demonstrate how to use information as entanglement to integrate, control, automate, and optimize heterogeneous processes that span systems of fundamentally different natures (humans, machines, smart agents as in the case of Human Machine Teaming).
We thus seek research and partnership opportunities to expand and mature the concept and application of the IDT/HDT.
A typical output of such research is an IDT/HDT prototype that automates and optimizes heterogeneous processes based on an information model (the Information Heat Map, IHM) that captures the correlations between the parameters of the various systems, notwithstanding their nature.
Realizing General Intelligent Solutions: The Information Digital Twin (IDT) & The Social AI Platform
The Information or Human Digital Twins (IDT/HDT) are the technical implementation of the unified information and intelligence concepts. The IDT/HDT architecture and algorithms enable them to assist their users in measuring and assessing the level of information (entanglement) involved in their interactions with their environment and help them optimize their actions towards increasing their information (i.e., entanglement) over time.
Entanglement, Predictability & Information
Our assumptions concerning the unified definitions of information and intelligence are founded on the concept of entanglement, a fundamental characteristic in quantum mechanics. According to Erwin Schrödinger, 1948, "entanglement consists in the fact that a single observable (or set of commuting observables) of one system is uniquely determined by a single observable (or set of commuting observables) of the other"
For two entangled systems, say A and B, knowing the state of one of its parameters provides system A with complete knowledge (prediction!) of the corresponding parameter in system B. The higher the entanglement between the two systems, the higher their ability to predict and influence one another.
As we propose, information is thus an indication and a measure of the level of entanglement between two systems, which corresponds directly to their ability to predict and be predictable.
How the IDT uses Information to Enable Intelligence
As an AGI agent, the Information Digital Twin, IDT, or the human-specific version, the Human Digital Twin, HDT, enable their users to measure their level of information (entanglement or predictability) as they interact with their environment. This measurement is based on calculating the probabilities of the various parameters involved in the user-environment interactions as captured in the Information Heat Map (IHM). The various use cases (Shipment Digital Twin, Human-Machine Teaming) provide examples of the information map structure used to calculate the user-environment information.
As indicated in the two examples, any change of any involved probability would change the value of the user-environment information. The IDT can then use the information change to update its search for alternative actions to maintain or increase the user-environment information, which reflects the IDT intelligence.
Currently, we have two pending patents covering the Social AI Platform's (and the Information Digital Twin) novel techniques, architecture, and algorithms
Research Focus-Improving RL Agents Performance and Adaptability
Current RL (reinforcement learning) agents rely on an interpreter* (or the agent designer) to establish the task states, rewards associated with different states, and agent objectives (e.g., end state). Changes in the environment and/or the agent's objectives necessitate the interpreter to intervene and adjust the states and/or rewards to adjust the agent's behavior accordingly.
Suppose the agent can independently update its states and rewards based on its objective. In that case, it will become more adaptable to changes in the task environment and more general because it can adjust its behavior to multiple objectives.
Using Agent Information (its Level of Entanglement with the Environment!) to Optimize its Actions
Replacing the observer with the Information Heat Map (IHM), which captures agent-environment interaction information, will allow the agent to dynamically update its states and rewards based on its objectives and thus be able to adapt to changes in the environment.
The current emphasis of our research is to validate this assumption. Furthermore, suppose we assume that the agent's ultimate objective is to improve its total information. In that case, the agent can use changes in interaction information to simulate numerous actions and choose the ones that will increase its information in the long run.