The basic idea is simple. Instead of having a single control unit sequencing the operations of the machine in series (except for certain subsidiary operations as certain input and output functions) as is now done, the idea is to decentralize control with several different control units capable of directing various simultaneous operations and interrelating them when appropriate…
A parallel control machine could locate points in itself needing repair. It would tend not to be completely incapacitated by any single material failure…
This does not require special design of the machine to take instructions more readily but rather the development of interpretative programs by which translation from more abstract into more explicit instructions can be effected by the machine itself… Here parallel control offers a distinct advantage since it permits the translation to occur simultaneously with the computation, indeed several different stages of translation might be occurring simultaneously…
If logical units could be made cheap enough, and the same for memory, one could afford to have trial and error and search processes used in the machine. Trial and error processes and learning processes (which would require a lot of memory) would be helpful in developing high order interpretative capacity. Ultimately trial and error processes, combined with search or association processes, abstraction processes, and learning or conditioning processes should lead to the learning machine or even to the genuine thinking machine. For this development parallel logical operation will certainly be important.
That was John Nash’s excellent, and highly readable, RAND research paper: Parallel Control (1954). A lot of the logic behind that paper, and the logic of distributed systems, should also be considered in the context of complex social systems like military acquisition. It is precisely the reason why we like markets as well as democratic governance.
Leave a Reply