An Experimental Framework for Intelligent Systems

Figure 2.8: Evolutionary Framework For Genes and MEMs

Making things still more concrete we have to define an experimental setup for the realization of the computational framework.Prior to computation we distinguish at least three different types of evolutionary processes (for the genetic and memetic evolution cf. figure 2.8:

1. Chemical: The evolution from basic atoms through bio-molecules to the first self-reproducing biological structures known as 'cells'.
2. Genetic: The evolution from first cells to multicellular structures known es plants and animals including homo sapiens, whose bodies are individually generated ('ontogenesis', 'growth') based on the genetic blueprint of there genomes
3. Memetic: the evolution of knowledge from basic memory contents to complex conceptual structures manifested in knowledge artifacts.

Although the genetic and the memetic evolutions are completely different, structurally they are similar. While in the genetic evolution the agents are cell structures which act as individual systems of a population which can exchange genetic information for the production of a new system based on these recombined genetic information we have in a memetic evolution agents of a population which not only can exchange genetic informations but also memetic information by symbolic communication. These memetic information can be mediated by an internal memory system as well as by external memory systems. The memory system as well as the communication process can operate on the memetic information and change it. Usually memetic information encodes parts of the world experience of an agent associated - and embedded - in symbolic structures.

For the goal of intelligent systems supporting humans in a human like way we are primarily interested in the memetic evolution. In the ideal case we would repeat the chemical and the genetic evolution to generate base systems for the memetic evolution. But the feasibility of such simulations is completely outside of scope (compared to the chemical and genetic evolution the simulation of the physical and chemical process of the known universe with regard to the development of the stars is comparably 'simple'). Nevertheless there is an approach to try to 'compromise' between not tackling with the complexity of a full genetic evolution simulation and the need for minimal evolutionary structures providing the base system. The idea is to structure the genetic information as a formal grammar which describes complete agents (cf. for the idea of combining GAs with grammars e.g. Antonisse 1991 [7], Vonk et al. 1995 [333], Hickingbotham et. 2008 [133], Sheikholharam et. 2008 [281], Arabsorkhi et. 2009 [8], Choubey et. 2010 [49], Luna et. 2010 [198], Pandey 2010 [241]).

The primary idea for the genetic evolution to evolve successful agent structures is to use a formal grammar - e.g. a context free grammar (CFG) - to specify a set of possible agents. The initial population contains many different such grammars. One adapts the genetic operators to these structures and starts the genetic process as usual. For the feedback one can define a multi-objective feedback function rating different aspects of the performance in a combined evaluation yielding a set of feedback values which can be ordered. The feedback cycle will usually not be identical with the simulation cycle because success or non-success can only be computed after longer action sequences.

From the genetic evolution one has to distinguish the memetic evolution. The memetic evolution happens inside and between living agents. Using their memory-function they will extract information from the world and the agent bodies, will generate abstractions, rules etc. and thereby they will set up some knowledge.