Ask HN: Bayesian FTW
3 comments
I honestly don't believe Bayes is that relevant to day to day decision making but I'm keen to see a counter-argument (I have some but I don't find them convincing yet).
I see Bayes as a method to solve certain classes of problem but the key challenge in everyday decisions is more like a design problem i.e. to define what the problem is. Once you have defined "what would be a good outcome of this decision" sufficiently, the answer rarely requires statistical methods. The common sources of error in a decision are in its definition e.g. omitting a requirement that is later revealed as essential.
For example, if you are analysing drug studies, Bayes is obviously relevant but for something more common place such as choosing a software tool, the main challenge is to understand what your goals are so that you can identify the criteria by which to judge the tools and the trade-offs you are willing to accept.
The biggest steps to improving decision quality appear to be process related e.g. using prototypes to explore options before making a larger commitment. Such acts are so effective because they reveal information that lets you improve your goals rather than just clarifying the quality of an option.
I see Bayes as a method to solve certain classes of problem but the key challenge in everyday decisions is more like a design problem i.e. to define what the problem is. Once you have defined "what would be a good outcome of this decision" sufficiently, the answer rarely requires statistical methods. The common sources of error in a decision are in its definition e.g. omitting a requirement that is later revealed as essential.
For example, if you are analysing drug studies, Bayes is obviously relevant but for something more common place such as choosing a software tool, the main challenge is to understand what your goals are so that you can identify the criteria by which to judge the tools and the trade-offs you are willing to accept.
The biggest steps to improving decision quality appear to be process related e.g. using prototypes to explore options before making a larger commitment. Such acts are so effective because they reveal information that lets you improve your goals rather than just clarifying the quality of an option.
Well, I wouldn't be surprised if the only practical benefit in building a Bayesian Network model ended up being that "that you had to think long and hard at what the various inputs are and how they should be linked".
I am curious if there is any kind of resource that would help me to confidently add Bayes to my normal decision-making toolchest.
I am curious if there is any kind of resource that would help me to confidently add Bayes to my normal decision-making toolchest.
Yep, I can see the value as a thinking tool. There are tools like http://www.bayesia.com/ and http://www.hugin.com/ but they are enterprise pricey and I find them very clunky.
I'd love to know how your ideal tool might work. Do you mind if I drop you an email (from your profile info)?
I'm working on a decision making tool to let you quickly describe reasoning with simple bullet point lists. It aims to be as quick as a making a rough note so that with just a few different types of bullet, you can create belief networks and computable decision models.
The tool is intended for a general audience so it currently uses simple weighted evidence. If you can tell me how you would like to work with probabilities then you might get the tool you want :)
I'd love to know how your ideal tool might work. Do you mind if I drop you an email (from your profile info)?
I'm working on a decision making tool to let you quickly describe reasoning with simple bullet point lists. It aims to be as quick as a making a rough note so that with just a few different types of bullet, you can create belief networks and computable decision models.
The tool is intended for a general audience so it currently uses simple weighted evidence. If you can tell me how you would like to work with probabilities then you might get the tool you want :)
Sorry, seen this only now... please feel free to write directly, thanks.
I find that having a bayes classifier handy can help a lot with categorizing documents or text. I manually keep track of my bank accounts using a spreadsheet, and I started using a classifier to automatically categorize the transaction. It's quite a time saver.
Can you explain what you use as categorization elements?
You train your own set by posting the document and the classification / category to the service. The training set is then used to calculate the prior probabilities needed by Bayes theorem. Once you have a proper training set, you can post a document to the service and it will return a list of probable categories from the training set.
Your set might look something like this:
Category - Document
negative - I don't like ice cream
positive - That's an awesome idea
neutral - The wind is blowing today
positive - We won!
You'll get a categorization of either positive, negative or neutral.
Your set might look something like this:
Category - Document
negative - I don't like ice cream
positive - That's an awesome idea
neutral - The wind is blowing today
positive - We won!
You'll get a categorization of either positive, negative or neutral.
Yes, well - I understand the general theory, but I don't understand how it applies to managing your bank account.
I do manage my bank account, but I don't see any need for an "classifier": when I input (or check) a supermarket bill I don't need much help in setting it to "GROCERIES". Basically anything that gets into my accounting files it's either something I know already what is about, or else something I need to investigate (e.g.: a speed ticket from a foreign country, routed to my credit card by the car rental company).
I do manage my bank account, but I don't see any need for an "classifier": when I input (or check) a supermarket bill I don't need much help in setting it to "GROCERIES". Basically anything that gets into my accounting files it's either something I know already what is about, or else something I need to investigate (e.g.: a speed ticket from a foreign country, routed to my credit card by the car rental company).
I've had this one stacked up on my reading list for a while:
http://camdavidsonpilon.github.io/Probabilistic-Programming-...
It details how to use the PyMC library via iPython notebooks. I'm not sure if iPython qualifies as an easy to use app though.
It details how to use the PyMC library via iPython notebooks. I'm not sure if iPython qualifies as an easy to use app though.
What I am looking for is something that could help me making practical use of Bayes in my day-to-day life (professional and/or personal).
So I am basically thinking of:
- books or articles detailing practical examples of how to apply Bayesian models to day-to-day choices (if these covered debug and testing activities it would be a big plus) - an easy-to-use app (desktop or tablet/smartphone) to build and play with Bayesian networks models.
Of course, if you have anything else to suggest along the same lines, please do.