Events are statements with time-dependent truth-values. Event-related terms and concepts have time-dependent meaning. [Related problem: "the new riddle of induction"]
To "reason about time" and to "reason in time" are different — reasoning takes time, and time changes. NARS gets a "personal sense of time" using its internal working cycle as a clock. The "real-time experience" of NARS measures the time interval between perceived events.
Temporal relations are defined relatively between events. There are two basic temporal relation: "before–after" and "when". Here NAL allows different granularity and accuracy in representation. Temporal relations are combined with logical connectors and copulas: two versions of conjunctions and 3 versions of implication (equivalence).
Tense: event compared to "now". Problem: "now" is a moving reference. Tense is used only in external communication, not in internal representation. The need to attach a time-stamp to each sentence to remember is creation (perceiving or deriving time).
Multiple ways to represent temporal information: (1) built-in temporal relations, (2) internal clock and tense, (3) acquired temporal relations.
In IL-7, the logical and temporal relations are processed in parallel by the inference rules. Then the inference rules are extended into NAL-7 strong rules, and their reversions becomes NAL-7 weak rules. Tensed sentences are processed similarly. NARS also use tensed judgment as evidence for eternal judgment.
An operation is a "realizable event" or "executable term", with an operator (as a relation), plus some input and output arguments (as a product). In this way, declarative knowledge and procedural knowledge are unified using "procedural interpretation", as in logic programming.
In programming languages, operations correspond to instructions, statements, routines, procedures, functions, etc. In the definition of an operation, an input argument can be an independent variable, and an output argument is a dependent variable. When executed, all input arguments are instantiated by constants, and the output arguments are determined accordingly.
Knowledge about an operation can either be inheritance/similarity or implication/equivalence. The inference rules are used as usual. Since NAL does not attempt to completely specify the conditions and consequences of an operation, it does not suffer from the Frame Problem.
Compound operations can be formed as compound events, and executed without much reasoning. With a given basic operation set, Narsese can be used as a programming language, and reasoning carries out planning, skill learning, self-programming, etc., like a logical programming language with the inference engine as an interpreter. NARS can also learn to program in another programming language by using Narsese as the meta-language.
The execution of an atomic operation is normally not the duty of the inference engine, but that of the host system (as the "body") or a peripheral device (as a "tool"). In the sensorimotor interface of NARS allows operations to be registered in NARS, with initial knowledge. In this way, NARS can serve as a mind of a robot or an intelligent operating system controlling various application software.
An "NARS+" is a NARS plus certain tools/organs.
Operations decide what the system can do, but not necessarily what the system will do. A goal is an event that the system desires to realize.
Traditional representations of "goal" in AI include state and utility/fitness/reward function. Their common problem is to represent a goal as certain, static, and complete.
AIKR implies that a system may have conflicting, competing, and changing goals. To handle this situation, a relative "degree of desire" is need. A desire-value is defined on every event, indicating the extent it implies a (virtual) desired event. Therefore, desire-value is a derivative of truth-value, and can be processed accordingly.
Goal processing in NARS:
If a goal matches an existing judgment, it has been realized to the extent specified by the truth-value of the later.
If a goal is an operation, to realize it means to execute the corresponding procedure.
Otherwise, to derive goals by using the inference rules backwards.
An input or derived goal is pre-processed via revision to modify the desire-value of the corresponding event.
The decision-making rule check the plausibility of a highly desired event, and turn it into a goal that is actually pursued. This is where binary decision is needed, in spite of the multi-valued truth-values, desire-values, and priority-values. Due to AIKR, goal alienation becomes inevitable.
[Compared with traditional decision theory, such as the expected utility hypothesis, NARS does not assume given probability and utility functions.]
[Compared with traditional AI techniques:
NAL-9 allows certain internal operations of NARS to be controlled by its own inference process.
The selection of operations is made by balancing flexibility, efficiency, and stability. A candidate list is given in Section 13.1 of the textbook.
After the full implementation of NAL-9, the control mechanism of NARS will be a combination of some "autonomic" processes and some "voluntary" processes.
Some mental operations will correspond to feeling and emotion, which are based on evaluation of various factors from the system's viewpoint (as extended desire-values) and the operations that are invoked as quick responses to the evaluation results. Emotion is necessary for a complicated intelligent system.
The system's internal experience corresponding to the mental operations are on the conceptual-level, and it forms the system's consciousness.
Like its "external" experience, the "internal" experience of NARS is also subjective, discontinuous, and incomplete. Since these two types of experience are produced by different operations, there is an "explanation gap."
NARS provides a unified mechanism to reproduce various cognitive functions.
For example, some cognitive functions can be uniformly represented as question-answering:
? --> T : searching
T --> ? : recognition
T <->? : naming
? ==> T : explanation
T ==> ? : prediction
T <=>? : summarizing
The relation between functions and inference types is not one-to-one, but many-to-many. In NARS, causal relation is an acquired notion and has domain-dependent meaning. There are also different types of "explanation". [Temporal induction and classical conditioning are closely related.]
[The following association is proposed by Peirce:
deduction : demonstration
induction : generalization
abduction : explanation
However, in general a cognitive function is provided in multiple ways by the cooperation of various inference rules.]
Similarly, many traditionally separated processes are absorbed into the reasoning framework:
perception as action,
reaction and planning,
incremental skill learning.
Compared with the various cognitive architectures, the unified approach followed by NARS has the advantages of consistency and simplicity.