Course 03-05-H-711.02

Cognitive Systems 2

Thomas Barkowsky, Sven Bertel, Denise Peters & Christian Freksa

Winter 2008/09
Mondays 10:00 - 12:00h & 13:00 - 15:00h, Cartesium Rotunde
4 SWS (ECTS: 6)

Exercise 1

One question developers of navigational assistance solutions may be interested in is the following:
Is memory for an unknown outdoor environment (e.g., some part of a city) better after exploring it with or without a map?

Design a psychological investigation, adhering to the six rules introduced in the lecture, to answer this question.

A proper description of the proposed psychological investigation will include detailed descriptions of:

Exercise 2

On the basis of the Eaters world introduced in part 2 of the Soar tutorial, create the following special eaters:

a) The bonus-food-despiser eater: this eater does not like bonus food and never eats it. When it is facing bonus food, it either tries to bypass the bonus food (preferred choice) or (if there is no normal food around) it jumps over the bonus food to get more normal food.

b) The smart bonus-food-lover eater: it tries to eat as little normal food as possible by exploiting the vertical bonus food lines and by making use of the fact that bonus food is available in regular vertical lines with known fixed distances between each other.

c) The lazy jump-over-the-wall-robber eater: this guy makes as few moves as possible and employs two alternating strategies: (1) it places itself next to a wall; (2) when another eater passes behind the wall and the score of the other eater is higher than its own score, then the eater jumps onto the other one which (as described on the first page of Part II of the tutorial) results in their scores being averaged.

Exercise 3

Extend the sentence verification system presented in the 4th unit of the ACT-R tutorial:

a) Such that the system properly understands and encodes sentences presented in passive voice.

b) Such that it is able to respond to the test sentences provided by the experiment system. The cognitive model should be able to compare a given test sentence with the corresponding facts stored in memory and to respond by pressing the appropriate key.

Hints how to address both tasks are given in the tutorial unit.

Exercise 4

A)
First install JavaNNS:

http://www-ra.informatik.uni-tuebingen.de/software/JavaNNS

B)
Get it to run and open the Error Graph Window and the Log Window, to see what you are training.
Open the "spirals.pat" from the "examples" folder and generate a feedforward network with the following layer setup:

2-5-5-5-2 (these are the layers)

Train the network (for this you will have to find the appropriate parameters and should use only 10.000 steps for training).
Use all three rules as learning rules, one after the other:
1.) Backpropagation
2.) Hebbian
C)
For the discussion of the result you need to explain each learning rule in detail: Provide the learning equation, a thorough discussion of the theoretical backgrounds of the learning rules (origins, design rationales, how do they compare to one another?). Quote all literature and other sources that you use. You need to also submit protocols with chosen parameters and snapshots of the learning curves for all rules.

Exercise 5

Time to reflect now that most of the course lies behind you. Please describe your perspectives on cognitive modeling. The following questions are meant as an inspiration to start you off, so please go beyond them. What is cognitive modeling for? What is it not for? Which are the basic paradigms? How do they compare in their philosophies, methods, strengths and weaknesses, or available tools? Where are the chief problems of the cognitive modeling approaches that we discussed in the course? How do you think one can circumvent these problems? Where are different methods and tools needed? Which application scenarios are well suited for cognitive modeling approaches? Which ones aren't? How do the paradigms that you learned about tie in with standard paradigms of software and application development? Should any paradigms be changed and if so, how?




  Sven Bertel, Denise Peters,  Thomas Barkowsky / 09 Sep 08