Quiz¶
I solved by finding only combo which has B, C, H because they are different from status quo or normality.
David does several things
- Principle of coverage. ensures answer covers all signs of the symptoms
- Principle of Parsiomony. Single simple hypothesis over combination.
- various interactions could complicate
- used the term explanation. diagnosis should explain the data.
What happens if no hypothesis that does not cover all data or multiple hypothesis that cover all symptoms?
Defining Diagnosis¶
- When actual behaviour differs from expected one, there is a discrepancy and we diagnose.
- Car engine not turning on when key is inserted. - Rule based reasoning
- Computer getting heated. Memory tells fan was problem in the past. I fix the fan. - Case based reasoning
- Program not working as expected - explaining how model works to duck - Model based reasoning.
So we saw different domains where diagnosis is used. Also methods as rule, case, model based reasoning.
Data and Hypothesis Space¶
- Data can be specific to abstract
- Hypothesis can also be from specific to abstract.
- Hypothesis explains data. Eg. Car is not starting because carberetor is not working.
Too many data, hypothesis, mapping within and between data and hypothesis could complicate
So the process usually abstracts and focuses on only what matters. For eg, in case of a patient, only focus is on symptoms data. Ashok has fever of 110 deg C could be abstracted as Ashok has high fever.
- Abstracted data mapped to abstract hypothesis. Ashok has a viral infection.
- Then abstracted hypothesis is refined. Ashok suffers from flu or whatever.
- We want hypothesis to be as refined as possible.
- As shown above, Bottom up classification happens in data space, mapped and then top down classification happens in hypothesis space. This is called heuristic classification, another method like CBR, rule based, or model based reasoning.
Problems with Diagnosis as Classification¶
Inferences¶
- Given a rule and a cause, deduce the effect. Always true.
- Given a cause and an effect, induce a rule. May be true
- Given a rule and an effect, abduce a cause. May be true.
- Inferences might be combined. Scientist example. First abduce a cause, induce a rule and then use deduction to predict.
Criteria for Choosing a Hypothesis¶
- Principle of Coverage - Hypothesis must cover as much of the data as possible
- Principle of Parsimony - Smallest number of hypothesis ought to be used
- Some hypotheses may be more likely than others.
Usually the hypothesis is a trade off of all 3 above.
- This approach can be used for elsewhere also. For eg, Intelligence Analysis.
Quiz¶
- Note Elevated E and Reduced E cancels out each other, so all symptoms are covered. Principle of coverage.
- However I chose Thetadesis: To choose a single hypothesis, principle of parsimony. That was also allowed as correct answer. Its a trade off between principles.
There are some important points, its better to watch David's answer here. Note, how he also explains confidence choices, and then discarded them in favour of other principles.
from IPython.display import YouTubeVideo
YouTubeVideo('asWcZv98Ahc')
Completing the Process¶
- Note this is like classification, the abstrated hypothesis representing equivalence classes.
- Note how the cycle could continue from treatment (in case of medicine, else generally action), to more data points.
