15

1

TL;DR

Given a patent application describing a calculation using a maximum likelihood approach, and a piece of prior art that has the same goal but uses Monte Carlo simulation, is this sufficient to challenge non-obviousness?

The documents

I'm looking at patent application 20120226640, Behavior and information model to yield more accurate probability of successful outcome (a more readable PDF version can be downloaded from FreshPatents).

This is really a business method patent, so whether this would actually hold in light of Bilski etc. may be questionable, but I'm not focusing on this issue.

However, the described method very strongly reminded me of Evidence based scheduling (EBS). This description from 2007 by Joel Spolsky, while being "just a blog post", has been extensively circulated and studied, so my impression is that it passes the test for being valid prior art.

Summaries

Here are my (obviously non-authorative) summaries of the patent application and of EBS.

The patent application

This is about salespeople giving an estimate of how probable it is that a particular deal they're pursuing will eventually be closed ("have a successful outcome", in the words of the application).

The salesperson gives their estimate (the "user-reported probability"), and based on this (and previous knowledge), the computer gives a better estimate that's more realistic, based on previous estimates and eventual outcomes.

The application calls this the "user-believed probability" under the assumption that the report may be skewed (consciously or sub-conciously) by the salesperson, and the computer-estimated value more closely reflects the salesperson's actual feeling of how probable a successful deal really is.

Evidence based scheduling

This is about software developers giving an estimate of how long it takes to complete a certain feature. Based on this estimate and knowledge about the quality of the developer's previous estimates (compared to the corresponding eventual completion times), EBS gives a better estimate that's more realistic (in fact, it even more detailedly gives a distribution of how probabable it is that the feature will be complete by any given date).

My obviousness argument

The main difference between the two is the statistical method that is utilized to improve the estimate. The patent application mathematically calculates its result using a maximum likelihood approach, while EBS uses a Monte Carlo simulation. Monte Carlo essentially is a method to estimate a maximum likelihood (it actually does more than that, but I'll leave it at this).

Now, here's my question: If the difference between the two comes down to the statistical method used to obtain the result, does this pass the obviousness test? Quoting from the FAQ (which in turn quotes Wikipedia),

One of the main requirements of patentability is that the invention being patented is not obvious, meaning that a “person having ordinary skill in the art” would not know how to solve the problem at which the invention is directed by using exactly the same mechanism.

In this case, a "person having ordinary skill in the art" would probably a statistician. And interchanging two different well-known methods for getting similar results looks like a fairly obvious thing to do for a statistician.

The claim

I'm just focusing on claim 1 of the patent application. All other claims either explicitly or at least more or less implicitly depend on 1.

Full-bold blockquotes are from the claim; if you read only those, you have the full claim.

A method comprising: receiving, by a processor, a report indicating a user-reported probability of a successful outcome;

This corresponds to a computer receiving a software developer's estimate for a given task. Note that "the feature is completed by day X" is a "successful outcome".

estimating, by the processor, a behavior and information model based on the report,

The "behavior and information model" is described in a little more detail below; the "processor" in EBS is obviously just the computer that does the calculations.

the behavior and information model including a behavior model component having a bias parameter and a consistency parameter,

In EBS, these parameters are actually there, but they are somewhat combined in a single parameter: the developer's velocity history vector. The description (in section 2) does in fact use the term "consistent":

This common estimator has very consistent velocities, but they’re below 1.0. For example, {0.6, 0.5, 0.6, 0.6, 0.5, 0.6, 0.7, 0.6}

(emphasis in original). Compared to the so-called bad estimator, where the velocities are "all over the map", having the values close to each other is a sign of consistency. The bias paramter isn't explicitly called out in EBS, but it is hinted at:

Most estimators get the scale wrong but the relative estimates right.

The "scale" in this case is precisely the bias parameter; it's pretty much an average velocity.

the behavior and information model including an information model component having a first user-believed probability of a successful outcome and a second user-believed probability of a successful outcome,

This is extremely broad in claim 1, which doesn't actually give any details about those values. It gets a bit more specific in subsequent claims, but for the purposes of claim 1, it's just "we have these two values".

And since we established the "user-believed probability" to correspond to the estimated shipping date, these are in fact there in EBS. Not only two such values, but one for each round of the Monte Carlo simulation.

Since I'm going for obviousness via substituting statistical methods, it may be irrelevant that EBS has more than two, but it should be noted that the claim says "having", not "consisting of", so it's not even excluding the possibility of more than two.

where estimating the behavior and information model based on the report comprises employing a maximum likelihood technique that determines the bias parameter, the consistency parameter, the first user-believed probability of winning, and the second user-believed probability of winning that maximize a likelihood;

Here's the difference: EBS uses a Monte Carlo approach in which those values are more implicitly present on the way to yielding estimated shipping dates (or rather, their probabilites), while the patent application uses a maximum likelihood to explicitly calculate those parameters in order to eventually use them for yielding a result.

and, using, by the processor, the behavior and information model to yield a model-determined probability of a successful outcome that more accurately reflects a probability of a successful outcome than the user-reported probability of a successful outcome does.

In EBS, this corresponds to the computer using the developer's velocity history to yield shipping date estimates that more accuratly reflect the probabilities than the developer's estimate does.

Once again, my question

Can the different statistical-mathematical approach be considered an obvious step for a "person having ordinary skill in the art"?

If the "art" under consideration is the design of processes, then I would suggest examining the "Strategy" pattern of software design: http://en.wikipedia.org/wiki/Strategy_pattern.

– Zoe – 2012-09-20T17:40:56.193

4

A short answer to your question is that employing different statistical methods could be sufficient novelty - but it depends on whether such a switch in approach would be a natural move for "one of ordinary skill in the art." Is such a move somehow counter-intuitive, or even contra-indicated? Is the result unexpected?

More significantly, though, prior art analysis does not end with the first claim of a patent. Please see my answers at Prior art for using a camera in self-driving cars and Microsoft have submitted a patent for a whack to silence a phone ringer. How similar does prior art have to be? for some explanation of additional elements of the analysis.

But wouldn't existing prior art for the first claim (or a few out of many) at least a) make that part of the application unpatentable, and thus b) eventually only make those claims valid that are so specific that they're basically useless? – balpha – 2012-09-19T16:52:53.097

@balpha You are correct. Each "higher-level" claim that is invalidated reduces the scope of the patent, presumably reducing the community of possible infringers and thereby the overall value of the patent. – user96 – 2012-09-19T18:00:21.013

1

Any change from any original "methods and processes", which could include the statistical methods and algorithms, could warrant a completely new patent, in which case it wouldn't invalidate the original patent.

Here are two different approaches that hopefully show what I have in mind --

Consider the "Ladder of Powers" approach to fitting a curve to a data set. If a simple linear fit doesn't work out, perhaps a logarithmic transform, or exponential transform might. This is just basic statistical analysis and people go to college every day and learn this.

Now consider a different environment where you don't have the computational power to perform that analysis, and instead you come up with something that is "good enough". Perhaps you cache 20 values and do a simple linear interpolation. You have horrible residual values, much worse than in a more precise curve fit, but your errors are within some acceptable bounds. Because "good enough" is seldom taught by much of anyone, your new "good enough" approach may well rise to the level of patentability.

1

reading your profile and these two answers has me pondering patenting my dip detector [http://stackoverflow.com/questions/19269638] I made for a ChipKIT PIC32 as well as the concept of that test in and of itself. It feels like no one has ever cared about quality before me; or I'm the first fool aggro enough to do all that.

– Chris K – 2013-12-30T04:20:41.470