It is arbitrary to define artificial intelligence as a study of agents. It is further questionable to make that arbitrary determination of its definition when neither the inspiration of AI nor its past and likely future embodiments are so structured.
There is no biological evidence of such extreme compartmentalization as agents infer. Early brain investigators thought there may be distinct brain organs or brain wave patterns that make decisions, feel emotions, finds predators in a field of vision, or constructs sentences. Decomposition of function in this way originally eluded researchers and the predominant belief is that enormous overlap and dynamicity in geometric location of function is common across the system. The idea that a set of elements can be devised that can be assembled to create efficient, intelligent systems is an artifact of early physics and its application to mechanical engineering. Complex systems are not that easily reduced to a fixed set of components.
After threatening the veracity of the gods, we can with less bias consider the question of clustering, which may have been the end goal of the question.
Clustering algorithms in computing are like investment diversification in finance. Investments are financial decisions based on assurance and diversification is an investment technique to factor concurrent doubt into a larger investment strategy — a synthesis of the opposing ideas of confidence and doubt. Similarly, clustering is a way to spread across computing machinery the burden of the computation, which algorithms spread instead across time. They are opposing ideas about how to spread computing burden, but are used conjunctively in strategies empling clustering algorithms.
To understand this more deeply, clustering is one of the techniques in a list of them that comprise the slow reversal of a consequence of John von Neumann's famous paper outlining the central processing unit idea1. His argument was surely easy to follow by its intended reader, the United States Army Ordnance Department.
- National defense was augmented by increased digital defense system reliability.
- System reliability was augmented by increased mean time between failures of the most failure prone components, the vacuum tubes.
- The number of tubes that must function for the system to function is dependent on the number of circuits involved in each operation on data and control functions to set up and execute that operation.
Therefore, the reliability of the defense system will be maximized if the computer had no more than one processing unit and all computation was centralized. The decision was made to have one CPU per computer and the first popular programming languages had constructs of ordered sequences and iterating through them to simulate parallel computation. In LISP, the repetition paradigm was implemented as tail recursion with DEFUN, CONS, CAR, and CDR. In FORTRAN, the repetition paradigm was implemented with DIM, DO, and a parenthetic moving index or list of indices in nested loops. From the generalization of these, along with branching, the field of algorithms design formed.
Multiple cores, mutiprocessing, multithreading, FPUs, GPUs, clustering, and many other supercomputing strategies were a result of finding more reliable digital hardware designs than vacuum tubes and developing them as VLSI, compilation, and linking technical conventions.
- Wide scenario simulation
- Hardware parity support
- Junction reliability optimization
- Capacitance reduction
- Crossover reduction
- Reduced power dissipation
- Gate level temperature compensation
- Synchronous multi-period clocks
- Various forms of caching
- Register use optimization
- Removal of instruction redundancy
Algorithms are an artifact of taking what would be largely parallel operations in nature and splaying them out in time so that they could conform to the single CPU high level computer design. Much of the design changes and feature additions to computers, including multi-threading and clustering, have been a gradual evolution from the FOR LOOP construct to parallel operations beyond just the number of bits in a native address, data type, or machine instruction.
That what is delegated to what node in a cluster may be determined by a single algorithm running in one physical processing core is an interim condition. The popularity and effectiveness of thread pools, connection pools, and parallel and independent load balancing are indicators of how limited the idea of primarily algorithm driven delegation really is. It is much more efficient for each potential processing entity, in this case a process in a node in a cluster, possibly via a listening socket, grabs the incoming information and request from a queue and performs the work, including any subsequent queuing or synchronous action.
Under this wider perspective, the design of delegation of work within an extensible and possibly hot-provisionable set of computing resources is absolutely part of the realm of AI. In fact, this wider perspective would include John von Neumann's original rational argument and the subsequent buy-in from his defense department stakeholders as also part of AI development.
These concepts are independent of the current (and probably temporary) categories of network based machine learning techniques derived from multilayer Perceptrons. Using the current ternary paradigm, whether
- Data is labeled and some difference function is used to compare the results of a guess at the parameter set employed against a subset of examples with the labels of that subset
- Data is not labeled and some other criteria of fitness is used to compare the results with previous results
to determine the next guess in a convergent learning system
- Data is a real time sensory descripotion of an external system into which the learning system is deployed to follow a behavioral trend considered adaptive on the basis of a value or advantage function,
is not particularly closely related to the degree to which a system is serialized into algorithms with loops to simulate parallelism or parallelized to conform more closely to biology. The trend toward parallel computing will continue, fueled by the same speed and computing resource cost considerations that fueled the multi-core VLSI general processing, high performance computing platforms, and accelerated hardware for floating point arithmetic, video rendering, and now learning circuit implementations inspired by machine learning algorithms.
- First Draft of a Report on the EDVAC, John von Neuman, 1945 — Section 5.6: "Accelerating these arithmetical operations does therefore not seem necessary—at least not until we have become thoroughly and practically familiar with the use of very high speed devices of this
kind, ... Thus ... The device should be as simple as possible, that is, contain as few elements as possible. This can be achieved by never performing two operations simultaneously, if this would cause a significant increase in the number of elements required. The result will be that the device will work more reliably ..."