(448e) Programmable Process Structures, Generated from a Network and from Functional Meta-Prototypes

This paper is about a non-conventional modeling approach, implemented in declarative, logical programming. Considering the usual trade-off principle, it is well structured, but not effective, for the time being. It is an in-house used experimental tool, tried for dynamic simulation based analysis of multi-scale, hybrid processes in a broad range of applications, from cellular signaling based biosystems to agri-food and environmental process systems.

Motivation and background

This non-conventional approach is motivated by the tendency to apply chemical engineers' way of thinking and Process Systems Engineering methodologies for understanding, design and control of nano- and tera-scale processes in a multi-disciplinary domain. These hybrid, multiscale models may be more complex, than the specific mathematical apparatuses. However, the basic cognitive elements of these heterogeneous structures and functionalities may be mapped onto unified elements of an executable program code, directly, without their representation in any well-defined mathematical construct. Recently, based on this basic idea of Direct Computer Mapping, as well as on its early applications, a new framework has developed for generation and simulation of "programmable process structures".

Developed methodology

It is obvious, that the structure of processes is more sophisticated, than the networks in sense of network science. However, the case-specific set of local functionalities of the elementary building blocks of these structures are less complicated, than the systems of integral and partial differential equations. As an intermediate solution we suggest a special process structure that can be generated from a network and from two functional meta-prototypes. The meta-prototypes and the derived process structures are prepared for the semantically distinguished, but syntactically uniform representation of the "model specific conservation law based, additive" measures, and of the "over-writable" signals.

In this approach the usual network based formalization of process structures is extended to a special net of unified state and transition elements, containing dedicated input and output slots for additive measures and signals, as well as parameter slots for parameters. The actual structure is defined by standardized connecting elements, corresponding to increases and decreases of (extensive) measures, as well as to reading and overwriting of the intensive properties and other qualitative or quantitative signals.

In various applications many state and transition elements can usually be modelled with the same local programs. Accordingly, we distinguish the actual elements and the program defining prototypes. The prototype elements contain symbolic input, parameter and output variables, as well as a local program, using these variables and limited number of global data. During the simulation, the actual elements start with the initial conditions and parameters, and the output values are recalculated stepwise with the knowledge of input and parameter data, according to the associated local program prototype. The distinguished input and output slots for extensive / intensive data and for signals make the combined execution of the balance-based and rule-based functionalities possible.

The data flow between the elements is determined by the connections. Actually, the receiving side of the connections select the topology-determined output message(s) from the output slots of sending element, according to the reading operator. Next it delivers the message(s) to the input slot of receiving element according to the sending side and the respective writing operator.

Both the actual elements, as well as the receiving and sending sides of the connections may be specified by a spatial identifier, determining the compartment and/or the level and/or the scale of the given entity. Also both the elements and the connections have a timing that defines temporal constraints for the execution, supporting the combined event-driven and time-driven execution of models.

Summarizing the methodology, the executable process models are generated from the description of the (optionally multilevel or multiscale) process network and from two functional meta-prototypes. The case-specific, actual functional prototypes, containing the local models may be copied and edited from the general meta-prototypes. The actual state and transition elements are parameterised and initialised concerning their case-specific prototypes. The actual model elements are executed by their associated prototypes. This execution and the connection-based communication between the state and transition elements of the "programmed structure" are solved by a general purpose kernel program.

In the proposed model representation the detailed description of possibility (design) space, as well as the measurements and the evaluating functions of the optionally multiple objectives may be embedded in the state and transition elements. Accordingly, in identification and optimization tasks the robust simulator may easily be combined with external meta-heuristic optimizers.

Experimental implementation

The code of the recent implementation is written in declarative, logical Prolog language that is a limited subset of first order predicate calculus. Prolog makes the easy description of the local program prototypes possible. This feature, combined with the applied, restricted and standardized way of model description, supports the application of a general kernel for the quite different actual models. In addition, this declarative language offers a formal description by itself, which fosters also the possible future implementations.

The automatic generation of the programmable structures starts from the GraphML based graphical implementation of state and transition meta-prototypes, as well as from the description of process network. Generation means multiplying the meta-prototype elements, supplemented with the connections between them. (It is to be noted, that the user can also start with the GraphML editor, from a scratch.) The generated graphical models can be edited in any (actually yEd) GraphML editor. In this step the user makes copies from the meta-prototypes, next declares the case-specific programs. This is followed by initialization and parameterization of the actual elements. Next a Prolog program interprets the GraphML description into the facts and clauses, declaring the dynamic model. Finally the general purpose Prolog kernel executes the various modes of dynamic simulation. This means a cyclically repeated processing of state elements, state → transition connections, transition elements and transition → state connections.

Applications and experiences

Programmable structures have been applied for quite different processes, from cellular biosystems through technological and biotechnological process units to Recirculating Aquaculture Systems, as well as complex trans-sectorial agri-food and environmental processes.

The methodology supports the unified generation of reusable process models. The generated models are reconfigurable and extensible, as well as the models can be incremented from any "final" state. The robustness of dynamic simulation is based on the built-in control mechanisms, initiated by the automatic checking of feasibility bounds for the underlying "model specific conservation law based" measures.

The simulation models can be transformed into State Space Models, where the measurements (i.e. the input data) of the control come from the output variables of the model, while the interventions (i.e. the output data) of the control usually modify the parameters of the model.

The robust simulation fosters the dynamic model based (sub)optimization with various, externally implemented meta-heuristic methods (like Genetic Algorithm). The basic principle, in accordance with the engineering way of thinking, is that instead of searching for an exact global optimum with a simplified model, it might be better to generate multi-objective suboptimal solutions from a more detailed model.

The unified approach helps the model based comparison of the sustainability and resilience specific features of natural and human built processes. The keynote elements are the "model specific conservation law based" balances, the inherently local and rationally global process architectures, as well as the cooperative evaluation feedback between the functionally connected subsystems. These can be applied in the model based organization of sustainable process networks, consciously.

Planned further development

Recently, there are Big Data with limited predictive abilities, on the one hand, and large models with lack of appropriate parameters, on the other. We are to collaborate in building "tandem models" based on mutual, cooperative evaluation feedback between 'a priori' modelling and data mining. Hence the cognitive simulation model could be continuously identified and validated by the accumulating data, while data based reasoning could be extended stepwise by the capabilities of the continuously evolving dynamic simulator.

We plan collaboration to apply the same approach for the implementation of stochastic and fuzzy models.

In other collaborations we ought to search for the more effective software implementations of Direct Computer Mapping, using possible new programming languages and environments, as well as model-friendly no-SQL databases.

According to the former naive tests, this non-conventional modelling approach might be familiar with some methods of micro-granular parallelism, also in sense of software / hardware co-design, so we seek for collaboration also in this field.