Intelligent Systems: Architecture, Design, Control

Rating:
Summary: Clarifies many of the issues
Review: If a machine can act appropriately in an "uncertain" environment, i.e. if the machine is able to increase its probability of success via the execution of this action, then the machine should be designated as intelligent according to the authors of this book. The success of an action is defined as the achieving of subgoals that allow the machine to achieve its ultimate goal. The criteria for success and the ultimate goal are however defined by an observer or entity that is external to the machine. Thus an intelligent machine requires assessment from this observer in order to judge whether its actions are indeed successful. It is unable to make that assessment itself.

The authors spend a great deal of time elaborating on this notion of intelligence and how to actually incorporate it in real machines. Central to the authors' notion of intelligence is the ability of a machine to do searching and to engage in the formation of combinations, as well as the ability to group, cluster, or lump entities together into sets and to focus attention on the details of data or perceptions while ignoring the rest. The authors also assert that there are degrees or levels of intelligence that are determined by the computational power of the "brain" of the machine, the algorithms that are used by the machine, the information stored in its memory, and the sophistication of the processes that run the machine.

The authors also believe that the intelligence of a machine can also increase and evolve, due to increases in computational power and knowledge. A machine can become more intelligent only through learning through experiences. The authors are careful to note that learning is not required for a machine to be intelligent, but is required if the machine is to become more intelligent. They do not quantify though how much more intelligent a given machine can be as a result of learning. Can a machine become twice as intelligent after the process of learning? How about three times as intelligent? How exactly does the intelligence scale as the learning process is active?

An entire chapter of the book is spent on what actually constitutes learning in machines. The authors distinguish between `quantitative' or `parametrical' learning, and `cognitive' learning. The former is associated with adaptive systems where these systems can adjust to a changing environment. The latter is associated with a situation where the machine hesitates to make further generalizations or continue to "tune" itself. The authors list several axioms that they believe every learning system must satisfy. These axioms go by the names of representation, control, plant, goal, observables, and system. In a nutshell, they assert that any reality can be modeled, that a control system exists, that a system exists that is controlled, that a goal set and output specifications exists, that a set of observables can be measured or recorded, and that there is a basic "system level" description of the intelligent machine. In addition, a `learning control system' is introduced that is basically a control system that generates input according to a control law that is constantly redefined and recomputed.

To build an intelligent machine that will rival the performance of natural intelligences, the authors attempt to build a computational model of intelligence that consists of a closed loop of four fundamental processes. These processes are `behavior generation', `world modeling', `sensory processing', and `value judgment', and they work together so as to process sensory information, use and maintain knowledge bases, and to engage in goal-directed behavior. These processes are quite sophisticated and any intelligent machine must be capable of engaging in all of them, according to the authors. Indeed, their sensory processing function does not merely gather sensory information, but also detects and groups features, engages in object recognition, and compares observations with expectations. World modeling involves the construction of representations of various entities, events, and situations and also makes predictions, generates beliefs, and formulates estimates of the probable results of future actions. Value judgment computes the costs and benefits of various plans, along with performing risk analysis and calculating expected payoffs for these plans.

Their construction of intelligent systems is based on an architecture that organizes the joint functioning of devices that are not intelligent themselves. The architecture involves something called `functional closure', the construction of representations of the environment, learning via generalization, and algorithms that `self-reference.' An `elementary loop of functioning' (ELF) is used to organize a simple loop of activities involving the sensory processing (SP), the world model (WM), and behavior generation (BG) capabilities of the machine. This architecture is considered to be able to "perpetuate its own existence" and is therefore called a `generalized subsistence machine' by the authors. The fundamental property of an ELF is to exist as a goal-seeking machine, and ELFs can be part of other ELFs. Each ELF consists of a part that handles goal-directed functioning and a part that handles the regular subsistence functioning. The authors give a real-world example of this architecture, the NIST-RCS architecture, the name of which reflects the authors' host institution.

Particularly interesting is the authors' discussion of value judgments. Value judgments, according to the authors, provide criteria for making intelligent choices. This includes the cost and risk analysis of plans and actions, and the desirability of various objects. It is the view of the authors that machine intelligence cannot be achieved if the machine is unable to perform value judgments. This leads them into making definitions of emotions, priorities, and drives, and then to a notion of a `value state-variable', examples of which include goodness, pleasure, pain, hope, frustration, hate, fear, and confidence. The values for the state-variables are determined by the `value judgment functions' that reside in `value judgment modules' of the machine. A value judgment function takes an input state vector which describes the conditions in the world model and produces an output vector consisting of value state-variables.