User Tools

Site Tools


public:events:challenges5-team-results

<-main workshop page
humanobs header

Challenges 5: Team Conclusions & Solutions

Lecture: REQUIREMENTS FOR AGI
Author: Eric Nivel

Challenges

1. Can an AGI be hard-real time?
2. How to compress a program

Team A

Team Members:
Gudny R. Jonsdottir, IIIM, Iceland
Helgi Páll Helgason, Reykjavik U.
John-Jules Meyer, U. of Utrecht, the Netherlands
Ricardo Sanz, U. Madrid

Team Conclusions:

Prediction of time for underspecified problems is impossible. Assuming limited resources with unlimited potential for goals then the answer to the challenge is no. Under some optimal resource allocation, goals passed on to the system need to be theoretically achievable with available resources. System might need a “reject option” to be real-time, having the option of rejecting goals to ensure the timely completion of already accepted goals. Should reactiveness be sacrificed to guarantee realtime operation?

Black boxes in second challenge make almost it impossible to be real time. No longer possible to reason about its operation. If an AGI compresses its own programs, it may stop being an AGI (sacrifices generality/plasticity). Decompilation is not necessary, keep original parts around. They are slower but will be used less often. Compiled versions are more efficient but can not be changed.

Team B

Team Members:
Eric Baum, USA
Pei Wang, Temple U., USA
Mark Wernsdorfer, U. Bamberg, Germany
Yngvi Björnsson, Reykjavik U.

Team Conclusions:
determine time complexity restrict time complexity

trivial: restrict abstraction levels

hard-real-time?

trivial: no, because humans are agi, humans are not hard-real-time.
the environment is not completely predictable.
maybe its a bad idea to fix the available time to solve a task.
for classical algorithms hard-real-time can be determined,
for intelligent systems it defies adaptability

compressability?

trivial: compilation
lossless compression, because decompression may be neccessary
compilation as analogy to automatisation in humans
source code forgetting should not be immediate, but after a certain time, maybe gradual
humans do lossy and lossless compression
decompression for explanation and modification
gradual compression compilation (explicit and implicit/player and coach)

Team C

Team Members:
Hannes Högni Vilhjalmsson, Reykjavik U.
Deon Garrett, IIIM, Iceland
Haris Dindo, U. Palermo, Italy
Marjan Sirjani, Reykjavik U.

Team Conclusions: Real-time depends on the task, timing depends on compromises on what is important and what is not! The question of urgency is important: f.e. catching a ball thrown at you might not be important but it is urgent - otherwise it could hit you. Humans have compressed a lot of these procedures: we don't deliberate on many issues, we just execute them! How to decide what to compress and what not? An idea could be to have a measure based on urgency/importance at different scale (time, complexity, …). We could have models at different level of granularity as well (e.g. Fitt's law), and depending on the requirement we can chose a model as accurate as necessary. This can be handled in a winner-takes-all fashion, where different models compete for execution. We could have real-time systems outputting a decision every time one is needed, but it is not necessarily the optimal one. For example, we have reflexes which we cannot avoid, and are automatically activated. Are they compressed programs emerged during the evolution? We thing models should not be deleted when compressed into programs; the two would rather live side-by-side instead of being decompressed. Even if skills could interfere with each other (e.g. being an expert tennis player starting to learn how to play badminton).

Team D

Team Members:
Antonio Chella, U. Palermo, Italy
Hrafn Th. Thorisson, IIIM, Iceland
Eric Nivel, Reykjavik U.
Jörg Siekman, DFKI, Germany
Hamid Pourvatan, IIIM, Iceland

Team Conclusions:

Team D

Team Members:
Anna Ingolfsdottir, Reykjavik U.
Bas Steunebrink, IDSIA, Switzerland
Kristinn R. Thórisson, Reykjavik U. / IIIM, Iceland
James Bonaiuto, Cal Tech, USA

Team Conclusions:

public/events/challenges5-team-results.txt · Last modified: 2011/09/26 22:12 by thorisson