OpenCogPrime:GoalDrivenCognition
Contents
Goal-Oriented Cognition
Most of OCP's cognition is driven by goals. OCP's goal system is conceptually fairly straightforward, but technically quite complex topic, which interweaves all the other aspects of the OCP design.
We describe, in this chapter, OCP as an embodied, goal-driven, autonomous learning system. This material provides the conceptual and dynamical framework within which the cognitive mechanisms described in other pages of the OCP documentation operate.
The page OpenCogPrime:PredictiveAttraction reminds you of some necessary structures in temporal logic, important for interpreting some of the above-linked pages.
These cognitive mechanisms must be understood, in a OCP context, not as algorithms for solving isolated problems, but as components of a complex system achieving complex goals in complex environments. The interactions between the mechanisms will be discussed more in later chapters; here we will focus on their interoperation in the context of explicitly goal-seeking behavior.
XX insert from PPT basic diagram of world, perception, action, goals, feelings... XX
The above diagram loosely summarizes the sort of high-level control system that we believe is critical to intelligence. The parts of the diagram are generically meaningful and we propose that any intelligent system operating under limited resources must explicitly manifest these processes in some way. We will also indicate the types of OCP Atoms centrally or uniquely involved with each function.
Many but not all of the ideas in this chapter were implemented in the NCE in simplified form, in 2005, in the context of using the NCE to control a humanoid agent in the AGISim simulation world.
Goal, Feelings and Contexts
The pages supervened by this one largely involve three key concepts — Contexts, Goals and Feelings — and their interrelationships, and their (interdependent) roles in OCP system dynamics.
As it turns out, none of these concepts requires fundamentally new data structures. Rather, they may all be understood as wrappers around structures already existing in OCP for other purposes.
Goals do however involve a largely new process: the process of OpenCogPrime:EconomicGoalAndActionSelection, an outgrowth of economic attention allocation that is treated on another page.
Philosophically, we may say that, in the OCP conceptual framework, goals, feelings and contexts are represented as particular manifestations of deeper, more generic cognitive structures and processes.
Cognitive Processes Related to Contexts & Goals
Some unique cognitive processes associated with Goal and Context Atoms, reviewed in various of the pages supervened by this one, are:
- OpenCogPrime:ContextFormation — the creation of ContextLinks treating as Contexts states frequently experienced by OCP. This is done by a special instance of the OpenCogPrime:MapEncapsulationAndExpansion MindAgent, with parameter values tuned for context formation; and by the ConceptCreation and Clustering MindAgents.
- OpenCogPrime:GoalFormation — the creation of Goal Atoms representing goals that OCP wants to attain. The goal formation CIMDynamic, relying on PLN, seeks to form Goal Nodes that imply important and long-term-important Goal Atoms. Of course, this process is initially driven by the initial set of given Goal Atoms.
- OpenCogPrime:SchemaContextGoalTriad formation — carried out by PLN inference via a specialized control mechanism called predicate schematization, which is specially tuned to look for this kind of implication
- OpenCogPrime:GoalFulfillmentAndPredicateSchematization — the process by which a goal seeks schemata that it judges will cause its satisfaction, and activates them.
- OpenCogPrime:EconomicGoalAndActionSelection — the process by which some goals are chosen to be allowed to actively seek fulfillment, at a certain point in time; and actions are adaptively selected to fulfill these goals
Contexts/Feelings/Goals and Time
As with other Atoms, when a Context, Feeling or Goal Atom's TruthValue is evaluated at a certain time point, the value obtained is then merged in with the old value according to the revision rule. The temporal reweighting process is very important here, so that TruthValue reflects recent goal/context satisfaction more so than ancient satisfaction. A key point is that some Context Nodes and Goal Nodes must have a very high time decay factor, as compared to ordinary Predicates, because one sometimes wants the truth value of a Context Node to represent whether a context is relevant right now, not whether it was relevant last week.
Context, Goals and Importance
Oftentimes, part of a Goal or a Context will be a Link specifying that Atoms so and such are important in the system or Atoms so and such are important in interaction channel C. (A little later in this chapter, we'll review the notion of interaction-channel-dependent importance.) To achieve this the Goal or Context Atoms will contain Links of the form
HypotheticalLink (ExecutionLink Importance A x) HypotheticalLink (ExecutionLink (InteractionChannelImportance C) A x)
where A is an Atom and x is a number.