Download presentation
Presentation is loading. Please wait.
Published byKaren McDowell Modified over 8 years ago
1
Dr Matthew Berryman
2
Total systems intervention (Flood & Jackson): ◦ An umbrella framework for guiding the choice of systems methodologies (system dynamics, soft systems methodology, etc.) ◦ 3 phases: creativity, choice, implementation.
3
Knowledge-based expert system: ◦ Set of if-then rules. ◦ Easy for humans to read & follow ◦ Natural to break on distinguishing features.
4
Forward chaining: ◦ Start with the data available – details of the problem, and system – and work forwards to reach a conclusion – decision as to which method(s) to use. Backwards chaining: ◦ Start with a method, and work out what the problem & system would look like. ◦ If the expert system can’t identify a method, then pick the one that’s closest and work back.
5
Only as good as the expert(s). ◦ In terms of rules for distinguishing between the different methods. ◦ In terms of what methods are considered as outcomes. May be more than one for a completely specified set of data. Only as good as the user(s). ◦ Has the user correctly identified all the distinguishing characteristics? There may be multiple reasonable views of the system and hence multiple correct sets of distinguishing characteristics. ◦ Does the user follow it blindly (deliberately, or unknowingly)?
6
Based on whether the past can give an improved forecast of the future (causality can only go forwards). Stronger than just using correlation (avoids the sea level in Venice / bread price in the UK problem), but not 100% evidence for causality. Different statistical tests can be used: ◦ Original (regression based on asymptotic distribution theory) – can’t handle non-stationarity. ◦ Vector Error Correction Model (VECM). ◦ Vector AutoRegressive (VAR) model. ◦ Toda-Yamamoto modified Wald test.
7
Problems: ◦ Represent subjective beliefs. Assume fixed set of variables, and compute the probabilities. Can update the probabilities, but not the structure. ◦ Can’t have cycles (A→B→C→A). Image from: http://cli.vu/pubdirectory/67/huygen50.pnghttp://cli.vu/pubdirectory/67/huygen50.png
8
Benefits: ◦ Can handle cycles. ◦ Better training than BBNs. Problems: ◦ Can’t specify whether it’s A→B or B→A. Image from: http://homepages.inf.ed.ac.uk/rbf/CVonline/LOCAL_COPIES/AV0809/ORCHARD/
9
Doesn’t presuppose a causal structure, instead it infers one (the maximally predictive, minimal space) one from the data. Disadvantage: ◦ Applies to an output of a discretised (time and value dimensions) 1D time series data (x[k], x and k discrete). Some extensions of ideas to 2D CAs but they rely on the specific nature of CAs in constructing causal states.
10
if (you want to find causal relationships) { If (1D time series) { CSSR } else { Granger } } else (if you want to analyse a causal system with known relationships) { if (cycles) { Markov } else { BBN }
11
Adapt the decision tree. Fine tune the existing (exploitation, level 1) causal methods. Develop new ones (level 2). Proxies.
12
Despite limitations, I believe this to be a useful way of organising the set of causal methods we will research. High-level descriptions. Be adaptive!
Similar presentations
© 2024 SlidePlayer.com. Inc.
All rights reserved.