An Engineering Framework for Intelligent Systems

As has been stated before the primary goal of this lecture is the description how to built intelligent systems which can support humans in a sufficient human like manner. Therefore we have to translate the evolutionary framework into an engineering framework which does the job. As outlined in [87] and [88] (cf. figure 2.5 taken from [88]) a whole systems engineering management process can be highly sophisticated and complex.

Figure 2.5: Systems Engineering Management Overview

Nevertheless one can extract a strongly simplified structure out of the whole picture which gives the key elements of such an engineering process.

The process starts with a problem $\cal{P}$ of a stakeholder. Through a communication process the systems engineer translates $\cal{P}$ into a behavior model $\cal{M_{S-R}}$ that represents the complete expected behavior of the system to be designed:

$$requirements : \cal{P} \longrightarrow \cal{M_{S-R}}$$ (2.3)

Based on $\cal{M_{S-R}}$, the systems engineer develops during the design a system design model $\cal{M_{SYS}}$ which has to be verified:

$$design : \cal{M_{S-R}} \longrightarrow \cal{M_{SYS}}$$ (2.4)

$$verification : \cal{M_{S-R}} \times \cal{M_{SYS}} \longrightarrow \cal{V}$$ (2.5)

The design model $\cal{M_{SYS}}$ is further converted into an implemented system $\cal{M_{SYS*}}$ which again has to be validated. Validation is realized as a measurement process:

$$implementation : \cal{M_{SYS}} \longrightarrow \cal{M_{SYS*}}$$ (2.6)

$$validation : \cal{P} \times \cal{M_{S-R}} \times \cal{M_{SYS*}} \longmapsto \cal{V}$$ (2.7)

where $\cal{V}$ is a set of validation values indicating the correlation between the behavior model $\cal{M_{S-R}}$ and the system model $\cal{M_{SYS}}$.

With these assumptions one can construct a mapping from the environmental framework onto the engineering framework as follows (cf. figure 2.6):

Figure 2.6: Evolutionary Framework Mapped onto Engineering Framework

The problem introduced by the stakeholder in the engineering framework, is in the evolutionary framework implicitly given by the state of the world, which has to be mastered by a population. Te main task of a population in this world is to collect enough energy to enable it's members to exchange genetic information and to survive sufficiently often. As an indirect goal a population has to try to improve the genetic information to be able to adapt to changing conditions and to a growing competition with other populations. A first challenge is to find a first genetic description which is good enough to work as an initial code, which can be translated into some working structure. While the possible translation into a working structure is more a technical question which poses no serious theoretical problem the question of a first genetic information is rather challenging.

In the real evolution the question of the concrete conditions of the emergence of the first biological structures being able to use genetic information as a blueprint for a production of the real working structures is still not completely solved (cf. Eigen 1993 [83], Rauchfuß 2005 [249], Storch et al.2007 [313]:228f). We only know how self-reproduction works during evolution from that moment in history (somehow around -3.46 billion years from now), when the first working cells appeared on earth.

This leaves us with the challenge either (i) to define a mechanism which can eventually produce such an initial genetic description driven by an algorithm or (ii) to compile such an initial description manually. Case (i) is in principle possible because the known mechanism - a genetic algorithm (GA) - can do such a job. But without any kind of pre-knowledge this can become a very lengthy procedure. Thus in this lecture we try a combination out of (i) and (ii): in this lecture we try to get some initial knowledge about the problem space and then we define a genetic like algorithm to improve the proposed initial description.

From this the following subtasks are derived:

• Environment: One needs an environment which offers all those tasks which are assumed to be of relevance for the emerging population of intelligent systems. This environment can be a real environment or a virtual one. Usually we will only use such virtual environments which can structurally mapped onto real environments.
• Population: A population is a collection of individual systems sharing the same gene pool.
• Intelligent System: We assume that all populations under investigation have potentially intelligent systems. An intelligent system will be based on a genetic description which can be translated into a working system. Working systems are characterized by a 'departure' in time ('birth'), by 'growth', by 'mating', by 'aging', and by 'termination' ('death'). It is assumed further that intelligent systems possess internal structures which allow 'intelligent behavior'. While the term 'thinking' is a summary of genetically induced internal operations does the term 'learning' describe those kinds of interactions with the environment, which induce certain 'changes' inside of the system.
• Intelligence: To 'judge' about the fact whether a system can be classified as 'intelligent' clearly defined experiments have to be run. Depending from the results of these experiments a system is classified to have 'intelligence' to some measured degree.
• Performance: Besides the measuring of 'intelligence' many other parameters will be measured with the aid of controlled experiments.