Author: Alia
see also: Michiel's summary
Attendees: Alia, Michiel, Anna Tordai
In this course, Catholijn gives an interesting insight on her work on how to mimic animal (and human) learning behavior for modeling agents learning style. She discusses intelligent behavior, pattern behaviors, different types of behavior, and shared extended mind.
Additional notes:
Different types of learning mechanism (which can be modeled for agents):
Note: * requires some knowledge
What is the goal of understanding and mimic the learning mechanism: mind - extended mind - shared extended mind (as a collective)
Relation to my research: is there a way to model the learning behavior of people searching through the internet e.g. using parts of the "Information Foraging Theory" model from Peter Pirolli (see also his chi07 workshop paper: Towards a Tool for Predicting User Exploration ). If we can model and predict this behavior correctly then this should be the first step before trying to make adaptive content and interface.Martijn talks about how to model learning for agents based on Optimizing Control Theory, Operation research and planning research. Main idea: reward (and punishment) play for reinforcement learning.
Keywords (from the slides): Markov decision process, Bellman functions, Decision-theoretic planning (DTP), Reinforcement Learning (RL), Policy evaluation, Policy iteration, value iteration, exploration vs. exploitation, monte-carlo sampling, temporal-difference learning, Q-learning,
The more state spaces the environment, the more difficult it is to apply the exploration strategy.
Gerard gave an introduction to modeling argumentation. Reminds me of the Toulmin model a lot. One thing that sticks to my mind is that you can model an agent to pick a good argumentation strategy so that it can win a negotiation. (very much like the real life)
Relation to my research: Not sure. But it is nice to hear the lecture.
Bert discusses his work on qualitative reasoning applied in two domains: physics law and ecology. Even though qualitative reasoning will never replace numerical methods which is a vital tool for scientist there are several advantage of applying qualitative reasoning: 1.before applying numerical methods to get a feel of the result (more like quick and dirty prototyping), 2. as a tool for learning, 3. qualitative reasoning can be applied when the exact values (e.g. consonant) are unknown. 4. Bert also claims that qualitative reasoning can be applied as an argumentative tool.
I think the optimal solutions would be some kind of hybrid system e.g. to apply qualitative reasoning on top of numerical methods. So, scientist can still get the approximate value which they need but can do some kind of reasoning on top of it.
Implications to my research: can I use the GARP3 tool to model and predict user searching behavior?
Note: no slides available, copyright issues.
No further comment.
Relation to my research: Not very related.
Note: most well known IR model: boolean model, doc similarity, vector space model, probabilistic retrieval, language models, google pagerank.
Keyword: Relevant feedback algorithm by Rocchio.
Relevance to my research: Can I customize the probability ranking according to the different types of users information seeking needs?Keyword: topical relevance ranking
Introduction on question and answering retrieval. Interesting work, but as David pointed out there is still a lot of noise in solving the QA problem in the real WWW.