I am trying to build an Alexa Skills Kit, where a user can invoke an intent by saying something like
GetFriendLocation where is {Friend}
and for Alexa to recognize the variable friend I have to define all the possible values in LIST_OF_Friends file. But what if I do not know all the values for Friend and still would like to make a best match for ones present in some service that my app has access to.
Supposedly if you stick a small dictionary into a slot (you can put up to 50,000 samples), it becomes a "generic" slot and becomes very open to choosing anything, rather than what is given to it. In practice, I haven't had much luck with this.
It is a maxim in the field of Text To Speech that the more restrictive the vocabulary, the greater the accuracy. And, conversely, the greater the vocabulary, the lower the accuracy.
A system like VoiceXML (used mostly for telephone prompt software) has a very strict vocabulary, and generally performs well for the domains it has been tailored for.
A system like Watson TTS is completely open, but makes up for it's lack of accuracy by returning a confidence level for several different interpretations of the sounds. In short, it offloads much of the NLP work to you.
Amazon have, very deliberately, chosen a middle road for Alexa. Their intention model allows for more flexibility than VoiceXML, but is not as liberal as a dictation system. The result gives you pretty good options and pretty good quality.
Because of their decisions, they have a voice model where you have to declare, in advance, everything it can recognize. If you do so, you get consistent and good quality recognition. There are ways, as others have said, to "trick" it into supporting a "generic slot". However, by doing so, you are going outside their design and consistency and quality suffer.
As far as I know, I don't think you can dynamically add utterances for intents.
But for your specific question, there is a builtin slot call AMAZON.US_FIRST_NAME, which may be helpful.
My motivation for asking this question is that I have found an interesting problem using machine learning on a graph data set. There are papers out there on this subject. For example, "Learning from Labeled and Unlabeled Data on a Directed Graph" (Zhou, Huang, Scholkopf). However I do not have a background in artificial intelligence or machine learning so I would like to write a smaller program for a more general audience before working on anything scientific.
Several years ago I wrote a game called Solumns. It is an evil variant of the classic Sega game Columns. Inspired by bastet, it bruteforces for colour combinations disadvantageous to the player. It is hard.
I would like to improve upon its AI. I figure the game space (grid of coloured blocks, column positions, column colours) fits a graph structure better than a list of attributes. If that is the case, then this problem is similar to my research problem.
I am considering using the following high-level plan to solve this problem:
I'm thinking what would be useful is if the AI opponent could assign a fitness rating to a possible move based on more data than the number of existing squares on the board after the move. I'm thinking using a categoriser. Train on the move and all past moves, using the course of the rest of the game as a measure of success.
I am also thinking of developing a player bot that can beat the standard AI opponent. This could be useful when generating data for 1.
Use a sample of the player bot's games to build an AI that beats the strategic player. Maybe use this data for 1, too.
Write a fun AI that delegates to a possible combination of 1, 3, and the original AI, when appropriate, which I will determine using experimentation to find heuristic fudge factors.
To build the player bot, I figured I could use brute force to compute the sample space. Then use machine learning techniques such as those used in building Random Forests to create some kind of decision maker.
Building the AI opponent is where I am most perplexed.
Specific questions then:
Rating moves sounds like the kind of thing people do with chess, and although I'll admit my approach may be ignorant, there is a lot about this in literature and I can learn from that. Question is, should the player bot and AI opponent create the data sample? It sounds like I'm getting confused between different sample sets, which sounds like a recipe for bad training. Should I just play the game a bunch?
What kind of algorithm should I consider for training the player bot against the current AI?
What kind of algorithm should I consider for training an AI opponent against the player bot?
Extra info:
I am deliberately not asking if this strategy is a good fit for programming a game AI. Sure, you might be able to write a great AI more simply (after all it is already difficult to play). I want to learn about machine learning by doing something fun.
The original game was written in a mixture of racket and C. I am porting it to jruby for various reasons, likely with extensions or RPC calls to another faster language. I am not so interested in existing language-specific solutions here. I want to develop skills in this area and am not afraid to implement an algorithm for myself.
You can get the source for the original game here
I would not go for machine learning here. Look at game playing AIs.
You have an adversarial games (like Go) with two asymmetric players:
The user who places the pieces,
and the computer who chooses the pieces (instead of choosing pieces by chance).
I would probably start with Monte Carlo Tree Search.
I am doing a research on how grid explorer robots move.
most articles are about robots with no vision but with sensors to see state of all surrounding cells, but my robot has really no vision and only can sense a cell after it could or failed to explore it(imagine the sensors are 4 pushed buttons that will be pushed when it hits the obstacles.)
I was hoping someone could lead me to some clues of how to find papers about this kind of robots. I am really having a hard time figuring out how to find related information.
If you have a priori knowledge of the environment, but imperfect ability to sense your location in the environment (and probably imperfect actions) then what you're describing may be a partially observable markov decision process (POMDP.) Your four-button sensor fits nicely into that idea.
If you don't even have prior knowledge of the environment, then you need to augment the POMDP notion with elements of exploration or machine learning.
Note that POMDPs are geared toward scenarios with "rewards", and that while POMDPs are pretty well understood, efficient solutions are still a topic of research. As are ML/POMDP hybrids.
Environmental exploration and traversal with just push-button sensors for obstacle awareness, and without priori knowledge sounds to me like a Simultaneous localization and mapping (SLAM) problem.
If I recall correctly, Roombas use this, so there should be a fair bit of study done on it in that regard.
Locked. This question and its answers are locked because the question is off-topic but has historical significance. It is not currently accepting new answers or interactions.
How do you go about the requirements gathering phase? Does anyone have a good set of guidelines or tips to follow? What are some good questions to ask the stakeholders?
I am currently working on a new project and there are a lot of unknowns. I am in the process of coming up with a list of questions to ask the stakeholders. However I cant help but to feel that I am missing something or forgetting to ask a critical question.
You're almost certainly missing something. A lot of things, probably. Don't worry, it's ok. Even if you remembered everything and covered all the bases stakeholders aren't going to be able to give you very good, clear requirements without any point of reference. The best way to do this sort of thing is to get what you can from them now, then take that and give them something to react to. It can be a paper prototype, a mockup, version 0.1 of the software, whatever. Then they can start telling you what they really want.
See obligatory comic below...
In general, I try and get a feel for the business model my customer/client is trying to emulate with the application they want built. Are we building a glorified forms processor? Are we retrieving data from multiple sources in a single application to save time? Are we performing some kind of integration?
Once the general businesss model is established, I then move to the "must" and "must nots" for the application to dictate what data I can retrieve, who can perform what functions, etc.
Usually if you can get the customer to explain their model or workflow, you can move from there and find additional key questions.
The one question I always make sure to ask in some form or another is "What is the trickiest/most annoying thing you have to do when doing X. Typically the answer to that reveals the craziest business/data rule you'll have to implement.
Hope this helps!
Steve Yegge talks fun but there is money to be made in working out what other people's requirements are so i'd take his article with a pinch of salt.
Requirements gathering is incredibly tough because of the manner in which communication works. Its a four step process that is lossy in each step.
I have an idea in my head
I transform this into words and pictures
You interpret the pictures and words
You paint an image in your own mind of what my original idea was like
And humans fail miserably at this with worrying frequency through their adorable imperfections.
Agile does right in promoting iterative development. Getting early versions out to the client is important in identifying what features are most important (what ships in 0.1 - 0.5 ish), helps to keep you both on the right track in terms of how the application will work and quickly identifies the hidden features that you will miss.
The two main problem scenarios are the two ends of the scales:
Not having a freaking clue about what you are doing - get some domain experts
Having too many requirements - feature pit. - Question, cull (prioritise ;) ) features and use iterative development
Yegge does well in pointing out that domain experts are essential to produce good requirements because they know the business and have worked in it. They can help identify the core desire of the client and will help explain how their staff will use the system and what is important to the staff.
Alternatives and additions include trying to do the job yourself to get into the mindset or having a client staff member occasionally on-site, although the latter is unlikely to happen.
The feature pit is the other side, mostly full of failed government IT projects. Too much, too soon, not enough thought or application of realism (but what do you expect they have only about four years to make themselves feel important?). The aim here is to work out what the customer really wants.
As long as you work on getting the core components correct, efficient and bug-free clients usually remain tolerant of missing features that arrive in later shipments, as long as they eventually arrive. This is where iterative development really helps.
Remember to separate the client's ideas of what the program will be like and what they want the program to achieve.
Some clients can create confusion by communicating their requirements in the form of application features which may be poorly thought out or made redundant by much simpler functionality then they think they require. While I'm not advocating calling the client an idiot or not listening to them I feel that it is worth forever asking why they want a particular feature to get to its underlying purpose.
Remember that in either scenario it is of imperative importantance to root out the quickest path to fulfilling the customers core need and put you in a scenario where you are both profiting from the relationship.
Wow, where to start?
First, there is a set of knowledge someone should have to do analysis on some projects, but it really depends on what you are building for who. In other words, it makes a big difference if you are modifying an enterprise application for a Fortune 100 corporation, building an iPhone app, or adding functionality to a personal webpage.
Second, there are different kinds of requirements.
Objectives: What does the user want to accomplish?
Functional: What does the user need to do in order to reach their objective? (think steps to reach the objective/s)
Non-functional: What are the constraints your program needs to perform within? (think 10 vs 10k simultaneous users, growth, back-up, etc.)
Business rules: What dynamic constraints do you have to meet? (think calculations, definitions, legal concerns, etc.)
Third, the way to gather requirements most effectively, and then get feedback on them (which you will do, right?) is to use models. User cases and user stories are a model of what the user needs to do. Process models are another version of what needs to happen. System diagrams are just another model of how different parts of the program(s) interact. Good data modeling will define business concepts and show you the inputs, outputs, and changes that happen within your program. Models (and there are more than I listed) are really the key to the concern you list. A few good models will capture the needs and from models you can determine your requirements.
Fourth, get feedback. I know I mentioned this already, but you will not get everything right the first time, so get responses to what your customer wants.
As much as I appreciate requirements, and the models that drive them, users typically do not understand the ramifications of of all their requests. Constant communication with chances for review and feedback will give users a better understanding of what you are delivering. Further, they will refine their understanding based on what they see. Unless you're working for the government, iterations and / or prototypes are helpful.
First of all gather the requirements before you start coding. You can begin the design while you are gathering them depending on your project life cicle but you shouldn't ever start coding without them.
Requirements are a set of well written documents that protect both the client and yourself. Never forget that. If no requirement is present then it was not paid for (and thus it requires a formal change request), if it's present then it must be implemented and must work correctly.
Requirements must be testable. If a requirement cannot be tested then it isn't a requirement. That means something like, "The system "
Requirements must be concrete. That means stating "The system user interface shall be easy to use" is not a correct requirment.
In order to actually "gather" the requirements you need to first make sure you understand the businness model. The client will tell you what they want with its own words, it is your job to understand it and interpret it in the right context.
Make meetings with the client while you're developing the requirements. Describe them to the client with your own words and make sure you and the client have the same concept in the requirements.
Requirements require concise, testable example, but keep track of every other thing that comes up in the meetings, diagrams, doubts and try to mantain a record of every meeting.
If you can use an incremental life cycle, that will give you the ability to improve some bad gathered requirements.
You can never ask too many or "stupid" questions. The more questions you ask, the more answers you receive.
According to Steve Yegge that's the wrong question to ask. If you're gathering requirement it's already too late, your project is doomed.
High-level discussions about purpose, scope, limitations of operating environment, size, etc
Audition a single paragraph description of the system, hammer it out
Mock up UI
Formalize known requirements
Now iterate between 3 and 4 with more and more functional prototypes and more specs with more details. Write tests as you go. Do this until you have functional software and a complete, objective, testable requirements spec.
That's the dream. The reality is usually after a couple iterations everybody goes head-down and codes until there's a month left to test.
Gathering Business Requirements Are Bullshit - Steve Yegge
read the agile manifesto - working software is the only measurement for the success of a software project
get familiar with agile software practices - study Scrum , lean programming , xp etc - this will save you tremendous amount of time not only for the requirements gathering but also for the entire software development lifecycle
keep regular discussions with Customers and especially the future users and key-users
make sure you talk to the Persons understanding the problem domain - e.g. specialists in the field
Take small notes during the talks
After each CONVERSATION write an official requirement list and present it for approving. Later on it would be difficult to argue against all agreed documentation
make sure your Customers know approximately what are the approximate expenses in time and money for implementing "nice to have" requirements
make sure you label the requirements as "must have" , "should have" and "nice to have" from the very beginning, ensure Customers understand the differences between those types also
integrate all documents into the latest and final requirements analysis (or the current one for the iteration or whatever agile programming cycle you are using ... )
remember that requirements do change over the software life cycle , so gathering is one thing but managing and implementing another
KISS - keep it as simple as possible
study also the environment where the future system will reside - there are more and more technological restraints from legacy or surrounding systems , since the companies do not prefer to throw to the garbage the money they have invested for decades even if in our modern minds 20 years old code is garbage ...
Like most stages of the software development process its iteration works best.
First find out who your users are -- the XYZ dept,
Then find out where they fit into the organisation -- part of Z division,
Then find out what they do in general terms -- manage cash
Then in specific terms -- collect cash from tills, and check for till fraud.
Then you can start talking to them.
Ask what problem they want you want to solve -- you will get an answer like write a bamboozling system using OCR with shark technoligies.
Ignore that answer and ask some more questions to find out what the real problem is -- they cant read the till slips to reconcile the cash.
Agree a real solution with the users -- get a better ink ribbon supplier - or connect the electronic tills to the network and upload the logs to a central server.
Then agree in detail how they will measure the success of the project.
Then and only then propose and agree a detailed set of requirements.
I would suggest you to read Roger-Pressman's Software Engineering: A Practitioner's Approach
Before you go talking to the stakeholders/users/anyone be sure you will be able to put down the gathered information in a usefull and days-lasting way.
Use a sound-recorder if it is OK with the other person and the information is bulky.
If you heard something important and you need some reasonable time to write it down, you have two choices: ask the other person to wait a second, or say goodbye to that precious information. You wont remember it right, ask any neuro-scientist.
If you detect that a point need deeper review or that you need some document you just heard of, make sure you make a commitment with the other person to send that document or schedule another meeting with a more specific purpose. Never say "I'll remember to ask for that xls file" because in most cases you wont.
Not to long after the meeting, summarize all your notes, recordings and fresh thoughts. Just summarize it rigth. Create effective reminders for the commitments.
Again, just after the meeting, is the perfect time to understand why the gathering you just did was not as right as you thought at the end of the meeting. That's when you will be able to put down a lot of meaningful questions for another meeting.
I know the question was in the perspective of the pre-meeting, but please be aware that you can work on this matters before the meeting and end up with a much usefull, complete and quality gathering.
I've been using mind mapping (like a work breakdown structure) to help gather requirements and define the unknowns (the #1 project killer). Start at a high level and work your way down. You need to work with the sponsors, users and development team to ensure you get all the angles and don't miss anything. You can't be expected to know the entire scope of what they want without their involvement...you - as a project manager/BA - need to get them involved (most important part of the job).
There are some great ideas here already. Here are some requirements gathering principles that I always like to keep in mind:
Know the difference between the user and the customer.
The business owners that approve the shiny project are usually the customers. However, a devastating mistake is the tendency to confuse them as the user. The customer is usually the person that recognizes the need for your product, but the user is the person that will actually be using the solution (and will most likely complain later about a requirement your product did not meet).
Go to more than one person
Because we’re all human, and we tend to not remember every excruciating detail. You increase your likelihood of finding missed requirements as you talk to more people and cross-check.
Avoid specials
When a user asks for something very specific, be wary. Always question the biases and see if this will really make your product better.
Prototype
Don’t wait till launch to show what you have to the user. Do frequent prototypes (you can even call them beta versions) and get constant feedback throughout the development process. You’ll probably find more requirements as you do this.
I recently started using the concepts, standards and templates defined by the International Institute of Business Analysts organization (IIBA).
They have a pretty good BOK (Book of Knowledge) that can be downloaded from their website. They do also have a certificate.
Requirements Engineering is a bit of an art, there are lots of different ways to go about it, you really have to tailor it to your project and the stakeholders involved. A good place to start is with Requirements Engineering by Karl Wiegers:
http://www.amazon.com/Software-Requirements-Second-Pro-Best-Practices/dp/0735618798/ref=pd_bbs_sr_2?ie=UTF8&s=books&qid=1234910330&sr=8-2
and a requirements engineering process which may consist of a number of steps e.g.:
Elicitation - for the basis for discussion with the business
Analysis and Description - a technical description for the purpose of the developers
Elaboration, Clarification, Verification and Negotiation - further refinement of the requirements
Also, there are a number of ways of documenting the requirements (Use Cases, Prototypes, Specifications, Modelling Languages). Each have their advantages and disadvantages. For example prototypes are very good for elicitation of ideas from the business and discussion of ideas.
I generally find that writing a set of use cases and including wireframe prototypes works well to identify an initial set of requirements. From that point it's a continual process of working with technical people and business people to further clarify and elaborate on the requirements. Keeping track of what was initially agreed and tracking additional requirements are essential to avoid scope creep. Negotiation plays a bit part here also between the various parties as per the Broken Iron Triangle (http://www.ambysoft.com/essays/brokenTriangle.html).
IMO the most important first step is to set up a dictornary of domain-specific words. When your client says "order", what does he mean? Something he receives from his customers or something he sends to his suppliers? Or maybe both?
Find the keywords in the stakeholders' business, and let them explain those words until you comprehend their meaning in the process. Without that, you will have a hard time trying to understand the requirements.
i wrote a blog article about the approach i use:
http://pm4web.blogspot.com/2008/10/needs-analysis-for-business-websites.html
basically: questions to ask your client before building their website.
i should add this questionnaire sheet is only geared towards basic website builds - like a business web presence. totally different story if you are talking about web-based software. although some of it is still relavant (e.g. questions relating to look and feel).
LM
I prefer to keep my requirements gathering process as simple, direct and thorough as possible. You can download a sample document that I use as a template for my projects at this blog posting: http://allthingscs.blogspot.com/2011/03/documenting-software-architectural.html