Abstract

Nature has evolved an abundance of valuable and inspirational examples. It has even given birth to intelligent species. The concept of intelligence is valuable to us and deserves further exploration. The `brain processes' underlying intelligence are still little understood [Got97, p. 14]. Luckily, complex and often complicated phenomena have underlying principles that are not very complicated by themselves. In this thesis, (1) the principles for this evolutionary process were clarified and (2) evaluated for artificial application. (3) Finally we address whether intelligence can be said to emerge from it.

Ad 1) General principles thought to be essential are evolution, evolvable structures (substrates) and interaction with a rich and challenging environment. Specifically, neuronal structures have been essential to natural evolution of intelligence.

Ad 2) Both neuronal structures and evolution have been implemented artificially and have been combined, referred to as Artificial Neural Networks (ANNs) and Evolutionary Algorithms (EAs). Two experimental implementations are discussed and related to the theory. Evolution of virtual creatures' shapes to ANNs and EA and an artificial developing humanoid `baby robot' to developmental psychology. Implementation challenges and issues are discussed, scaling and interconnection problems. Possible solutions are use of FPGA, aVLSI, neuromorphic engineering, optic-holographic and molecular computing devices.

Ad 3) Do the implementations have what is needed for intelligence to emerge? Will intelligence eventually arise? The views of Roger Penrose and Alastair Channon [Cha96] are presented. Penrose conjectures that if the mind works non-algorithmically, we can not run it on any algorithmic machine or any Turing complete machine it is equivalent to. Channon argues that since we're unable to specify precicely what intelligence is, we should not expect it to emerge when using the `fitness function' in the traditional sense. Instead, Channon and others propose a co-evolution based aproach.