Set goal#
Add a single goal/idea to a “future” git graph, either wide or deep. Explicit backlinks:
Estimate cost#
Negotiate the goal(s)#
Unfortunately, you need to define the T in VNTE to quantify the V in VNTE. It’s the V that drives the generation of a T, but it can’t be quantified without it. Said another way, the V drives the search for any goal state, and the NT defines the set of goal states. If you aren’t specific about what you want, then you can’t be specific about how good it is (V). You can say you want more money, but if don’t define what you are going to learn to be able to do (specifically) to make money you can’t estimate how much money you will make. Said yet another way, V drives your imagination (of possible worlds). See also the NT in INVEST (mnemonic).
If the T is just given to you with intrinsic value (or trusted instrumental value), you can only imagine through subplans. See Estimate improvement value for comments on these terms.
Before any metrication or quantification, we must imagine possible worlds. These are theories in the context of the scientific method, and T in the context of INVEST. From this initial VT you can imagine a second VT (a variation on the first), or estimate an E to go with the first VT (to test the theory). Although SNI could come before E, even approximate numerical weight estimation (and therefore state space pruning) isn’t possible without E.
Goals should stay negotiable as long as possible. Keeping goals negotiable is the equivalent of putting ranges on targets, and using a logical disjunction (OR) as much as possible. It’s more work for the “authority” (buyer) to specify the details, but it can be the product much cheaper (faster, in software development). The less negotiable a goal is, the more innovation waste you’re going to see (less opportunity for coming up with 10 ideas to achieve X). Not only is this bad for the company, it’s bad for relationships (less control at work leads to less job enjoyment). Negotiability also prevent finger pointing about who should have understood or explained requirements.
Plans/stories should usually have only the what (NT) and the why (V), not the how (which is a sneaky way to extend the NT by claiming to provide an E). At the same time, suggestions should be welcome as long as they aren’t binding. What if your competitor has a good idea that isn’t IP protected? The product owner often does all the competitor research that no developer has time to do. Is anyone else doing competitor research? It is valuable. Why not use the research of others in academia? Isn’t even asking your coworkers for advice about how to complete something an example of giving up negotiability? Does being negotiable mean you get to decide how to do it alone?
In the language of Plan, imagine variations on the “specificity” of your goal. See also Types of Plans in Business: Breadth, Time-frame, Specificity & Frequency. How can you deliver some of the value without all of it?
Following this step, you should have several goals (e.g. B, K, L) with an estimate of the associated increase in value (in e.g. $) of achieving each goal. In theory, the value should be specified as a probabilistic function of the exact state achieved, but point estimates are a good starting point.
Said another way, it should be possible to produce the plan a..c as an alternative to a..b:
* b: G1, S2: 3PM
| * c: G1, S1: 2PM
|/
* a: 12PM
If it isn’t possible (e.g. G1 specifies the whole state of the universe at some point in the future) then it’s going to be harder to come up with e.g. 10 ways to solve a problem. If you don’t have flexibility in generating a solution (later) then a solution is going to be expensive or impossible.
A common message in planning in the context of Scrum is to reduce cost and value uncertainty at the same time. If you don’t you may find yourself with an extremely costly task (because the goal was too rigid) or an extremely unvaluable task (because the goal was too flexible).
Even if you’re targeting a focused goal there are likely many ways to achieve it (more than one “hot path” to it). Should you use a CNN or attention? See also Solution space. A more negotiable target gives you a larger solution space.
Not only are there multiple possible solutions in your solution space, for every solution there are multiple possible interpretations. For example:
The Bayesian vs. Frequentist interpretations of statistics
Interpreting a CONV layer as a FC layer (How are 1x1 CONV the same as a FC layer? - DSSE)
In general, prefer the term “focused” to “directed” research. The latter implies someone else is directing your attention, and we resist being controlled. The term “focus” is more neutral; it’s sometimes better to reduce our focus to a high-value subject. At other times, it’s better to expand our focus to attend to more state.
Objective function#
In the end, a value assigned to some goal state is a Function (mathematics) - Wikipedia where the cardinality of the domain and codomain are both one. The items in the domain are world states (e.g. including the assets assigned to everyone), and the items in the codomain are some measure of value (such as $). At best, everyone would provide a value function that assigns to every possible world state a $ value, so that we wouldn’t have to ask for the details of which particular “goals” they care about; we are all flexible and should be flexible with our goals.
Said another way, by default most people only consider a Partial function of world states to values. It would be better (in theory) if instead we all provided a complete Objective function. The product owner on a Scrum team (ideally) provides this objective function in the form of a person, so developers can query him/her with goal states.
If future value is uncertain (because of e.g. an uncertain future market), you may need to provide an Estimator of value rather than a non-probabilistic function. The “market” (in the present) provides an objective function to some degree in economics, though for specific subsets of world state and specific changes in world states. It can be expensive to query the market, more than asking a person, if market information isn’t readily available e.g. online.
Said another way, what makes a “variation” on one goal any different than a completely different goal? Is there some magical “distance” between the two goals that makes them variations on the same goal rather than completely separate goals?
We can’t compare models via their loss functions for the same reason we can’t compare people who disagree about what they regard as valuable. When comparing models, we often come up with some “evaluation” function that provides rigid objectivity in the presence of shifting values. It’s better to make datasets easy to visualize and model predictions easy to interpret, which allows for a changing market. It’s always possible to come up with at least one number that shows improvement, which you can come up with after you are convinced of improvement yourself.