• off-line architecture: in this case the computer is not connected directly to the process; the input of the data is performed either manually or through some movable information storage media, as for example magnetic disks, magnetic tapes, portable PC or specialised computers with relevant interfaces, etc. This type of architecture is currently mostly out of use due to the inconvenience and time delay necessary for data loading. However, in certain specific cases (e.g. collection of data from large amount of separated unconnected units) or in the case when preliminary human processing of the data is necessary (e.g. abstract or qualitative information should be input to the computer, especially if the data should be initially evaluated with respect to human expert specific, personal knowledge – e.g. taste, smell, outlook, colour or other hard to measure properties are to be evaluated) this kind of architecture can still be in use. Further, in case of simulation, post-processing analysis, and for the purpose of analysis, testing and research, the off-line architecture is popular thanks to its low cost and accessibility. Note that the off-line architecture is mostly applicable for relatively slow, usually man-controlled processes. An important indication for use of this kind of structure may consist in the fact that the necessary data sensors and measurement technology does not exist, is unavailable and/or relies on human or animal senses. Further, it may be the case that evaluation of selected product properties requires laboratory analyses, which both must be performed by man and require certain amount of time.

• in-line architecture: the computer is directly connected to the process, i.e. there is a physical flow of signals from the process to the supervising computer. However, there is no possibility to transmit information from the computer back to the process. This kind of architecture is typical for the process monitoring systems. It is usually applied at higher levels of two-level or multiple-level computer controlled systems. The data coming from the process to the computer is stored, processed and analysed; the output however is mostly directed towards human operators. The processed data may be used for various purposes, from warning and alarm generations, through parameter estimation and display to periodical report generation. This kind of architecture is mostly applied for processes where direct reaction of the system is either infeasible or unnecessary, where the relevant actuators are unavailable or do not exist, or where the reaction cannot be identified in a unique way and thus the final selection of the operation is left to the human operator (such situation can take place in Decision Support Systems (DSS)).

• on-line architecture: the computer is directly connected to the process; there is a physical flow of signals from the process to the supervising computer and simultaneously information is transmitted back from the computer to the process (and/or to the process controller). This kind of architecture is most popular in computerised process control systems, including direct digital control systems and hierarchical control systems. The possibility to transmit data back to the process allows for closed loop control at meta-level (e.g. for control algorithm selection or parameter modification), where the feedback phenomenon assures required performance in case of unexpected noise. In general, the on-line architecture is most powerful and demonstrates most interesting possibilities; the signal transformed back to process and/or its controller can be used for changing set points, control algorithm paratmetric or structural modification, activation and desactivation of certain controls, changing the mode and status, modification of process memory (if relevant), modification and adaptation of models used for optimal or adaptive control, etc.

The case of on-line control constitutes currently the most frequently applied type of architecture for control and becomes more and more in use for supervision. Two further notions, specific for on-line control are: real-time systems and any-time systems.

Real time systems constitute a specific class of on-line architectures. Although there exists no single, precise, widely accepted definition of such systems, most of the authors accept and put forward the following characteristics:
 

• on-line architecture: the system architecture is of on-line type, i.e. basing on the received from the process data the computer works out control and/or decisions which are directly applied to the process,

• time-critical requirements: the speed of the process requires fast reaction, i.e. there exists time requirements constraining the maximal time during which the control and/decision must be supplied back to the process; for specific situations this deadline time is defined a priori and can be denoted as t_d.

• guaranteed response time: it is assured that the computer system and operating algorithm is fast enough (both with respect to the hardware and the software -- applied algorithms), so that the control/decision will be always computed in apriori defined guaranteed response time or earlier; for specific situations this maximal response time can be denoted as t_r.

• satisfaction of time constraints: the maximal response time is less than the deadline time required while generating a response; for specific situations this condition can be denoted as 

In certain more complex systems, the deadline time t_d may vary depending on process state, external conditions, noise and current control goal. Due to very high speed of the modern computers and thanks to development of diversity of algorithms, it may turn out that a satisfactory solution is worked out much in advance with respect to required response time. In such a case the computer system can, using more data, memory, time and more advanced algorithms, attempt to still improve the solution. When finally the solution is ultimately required by controlled system, the last worked out, most optimal solution is returned.

Let  denote the minimal deadline time for response, i.e. surely no response is required before . Similarly, let  denote the maximal deadline for response, i.e. certainly, any answer must be provided before . Thus the first rough evaluation of the required response time is given by the interval . Note however, that contrary to the left border, the right border is not guaranteed. Let  denote the instant of time varying between and  and defining the current time instant when the answer becomes requested immediately. Further, let  denote the sequence of times necessary to compute the successive solutions, where  is the time for computing i-th solution,  and the i+1 – th solution is better than the i-th one, . For the system working correctly it is necessary to assume that , i.e. the first (reactive) solution must be available before the minimal deadline (as in the case of real-time systems). The further solutions corresponding to improved quality can be generated later on. When the request for solution is send to the computer (at ), the i-th solution generated at  such that  and  is returned to the system. Since in fact over the whole period when solution may be requested some solution is ready and waiting for the request, such system are named any-time systems. Further, it is reasonable to require that  since in the other case the best algorithm would never be in use.