Abstract:

Teaching concepts of cognitive science modeling is one aspect of advanced computer science studies in master level courses at University. In the last years, we have observed, that quite a lot of students have difficulties to transfer models, which have their roots in psychology, medicine and/or in biology, to computer science applications. Classical cognitive science, for example modeling in ACT-R (Anderson) is abstract in a formal way, and comparably easy to mediate. Neuronal networks and evolutionary algorithms are nor a more abstract level, and not so easy to grasp for students. In both cases, however, we observed difficulties in combining insights of Artificial Intelligence (AI) and Cognitive Science (CS) when it comes to modeling of so called „intelligent“ algorithms for certain applications.
Based on the idea of mini-worlds and explorative learning, we developed a software, which allows students to interactively apply their learned knowledge and to deepen their understanding of AI and CS concepts on a playful example.

The basic functionality of such an environment have to be:
First: Entities capable to decide and and act.
Second: Challenges that can be performed by these entities providing an assessable outcome.
Third: A simulation environment enabling perceptions and actions and set exercises to these entities.

We developed a „Happy Fish Tank“: the fish tank consists of a field with boundaries, randomly developed plants and stones, and inhabitants. The inhabitants are fish and kraken. To survive, fish need to find and eat food (which is generated on a random level) and kraken have to find and eat fish (we know that this is not taking place in real life).

The starter model and the starting simulation contains nothing more than this basic functionality - no rules or behaviors have been pre-implemented. The only concept we have integrated is that the entities (fish and kraken) try out behaviors and learn by success and failure. The task of the students is to develop models based on neuronal networks and cognitive science concepts (e.g. fish have to be „happy“, „healthy“ etc.), plus evolutionary algorithms, with the goal to develop a fish which is (after evolution) best adapted to the environment (which survive using goal-oriented mechanisms such as hiding, looking for food and other fishes and inherit their skills to the following generations). For this purpose, we have used the evolution concept NEAT (Neuroevolution for Augmenting Topologies), which is a genetic algorithm inspired by the mutation process of biological evolution.

We have accompanied the last winter term lecture „Cognitive Science“ and a programming course „Artificial Intelligence“ with project work in our „Happy fish tank“. As we only had 25 students, we have not been able to develop a significant quantitative evaluation, but we have made interviews with each student after the oral exams which take place after the course. We observed in the oral exams, that the students have developed comparably deeper rooted knowledge in AI and CS, compared to the years before (the course takes place annually). The grades reached in the oral exams were drastically better than before. As we have structured our interviews, we will repeat the experimental part of the course in this (and the next) year and see whether there is an innovation hype in our observations (which means, the grades have to decrease) or whether the basic concept proves to be good on the long run.

Keywords:

Genetic algorithms, neuronal networks, cognitive systems, neuroevolution of augmenting topologies.