Jie Cao | Dialogue, NLP, ML

Rethinking On-policy Optimization for Query Augmentation

2025-10-20T00:00:00+00:00

We are excited to announce our new preprint “Rethinking On-policy Optimization for Query Augmentation” is now available on arXiv!

About the Work

Recent advances in large language models (LLMs) have led to a surge of interest in query augmentation for information retrieval (IR). In this work, we present the first systematic comparison of prompting-based and RL-based query augmentation across diverse benchmarks, including evidence-seeking, ad hoc, and tool retrieval.

Key Findings

Our key finding is that simple, training-free query augmentation often performs on par with, or even surpasses, more expensive RL-based counterparts, especially when using powerful LLMs.

Novel Contribution: OPQE

Motivated by this discovery, we introduce a novel hybrid method, On-policy Pseudo-document Query Expansion (OPQE), which, instead of rewriting a query, the LLM policy learns to generate a pseudo-document that maximizes retrieval performance, thus merging the flexibility and generative structure of prompting with the targeted optimization of RL.

We show that OPQE outperforms both standalone prompting and RL-based rewriting, demonstrating that a synergistic approach yields the best results.

Authors

Zhichao Xu (First Author)
Shengyao Zhuang
Xueguang Ma
Bingsen Chen
Yijun Tian
Fengran Mo
Jie Cao (Corresponding Author)
Vivek Srikumar

Abstract

Recent advances in large language models (LLMs) have led to a surge of interest in query augmentation for information retrieval (IR). Two main approaches have emerged. The first prompts LLMs to generate answers or pseudo-documents that serve as new queries, relying purely on the model’s parametric knowledge or contextual information. The second applies reinforcement learning (RL) to fine-tune LLMs for query rewriting, directly optimizing retrieval metrics. While having respective advantages and limitations, the two approaches have not been compared under consistent experimental conditions. In this work, we present the first systematic comparison of prompting-based and RL-based query augmentation across diverse benchmarks, including evidence-seeking, ad hoc, and tool retrieval. Our key finding is that simple, training-free query augmentation often performs on par with, or even surpasses, more expensive RL-based counterparts, especially when using powerful LLMs. Motivated by this discovery, we introduce a novel hybrid method, On-policy Pseudo-document Query Expansion (OPQE), which, instead of rewriting a query, the LLM policy learns to generate a pseudo-document that maximizes retrieval performance, thus merging the flexibility and generative structure of prompting with the targeted optimization of RL. We show OPQE outperforms both standalone prompting and RL-based rewriting, demonstrating that a synergistic approach yields the best results. Our implementation is made available to facilitate reproducibility.

This preprint represents ongoing research in the intersection of large language models and information retrieval. We welcome feedback and collaboration opportunities.

shortcut collections

2016-02-20T12:09:00+00:00

Outline

Mac shortcut
Bash shortcut
Vim shortcut
IDEA shortcut
tmux shortcut

Mac shortcut

more details: https://support.apple.com/en-us/HT201237
The most important shortcut is the “Document shortcuts”
What’s more, those shortcut is globally usable in all the application in mac. it is recommended that please disable the arrow key, and change “caps lock” to “ctrl” key, which is also a style of Apple-II computer, now kept in HHKB keyboard or others.

Bash shortcut

more details: http://ss64.com/bash/syntax-keyboard.html
some alt key will not usable.
Most commonly used shortcut is also the document shortcuts in command line.
Beside that, the “history” part and “process control” part are all very important.

Vim shortcut

more details: http://bullium.com/support/vim.html
for vim, there are many cheat sheet online.
every one also has one’s own vim configuration file, then their own vim shortcut normally, we can support the above Mac short cut in vim.
Especailly for the “Home” and “End”
For support vim shortcut in other IDEA or applications, we always can find some vim extensions for this purpose. For example, “cVim” for chrome, “ideaVim” for IDEA.

IDEA shortcut

Beside mac shortcut, vim shortcut can be used in IDEA, there are many not common shortcut (only can used in IDEA) one start point for shortcut for IDEA should be “Command+Shift+A”, which can find the shortcut for every actions predefined in IDEA. Hence, extended with ideaVim emulator and this action find shortcut, we can quickly learn many keyboardless operation. It is very productive, since it combined the advance of using IDEA navigation operations with that of vim editor naviation shortcut.

tmux shortcut

tmux is something like screen with better screen and pane management. https://gist.github.com/MohamedAlaa/2961058
Pay attention to leader key, and how manage session, window, pane.

Note3-AofA: Recurrence

2015-09-19T16:31:00+00:00

The notes will introduce the types of recurrence, and the corresponding solution for all these recurrence, including some general telescoping methods and Master Theorem etc.

Telescoping a (linear first-order) recurrence

Linear first-order recurrence telescope to a sum.

When the coeffients are 1

Just telescope it. $a_n = a_n+n$ with $a_0=1$

Elementary discrete sums

* geometric series
* arthimetic series
* bionmial (upper)
* bionmial theorem
* Harmonic numbers
* vandermonde convolution 

When the coffeients are not 1

Multiply/divid by a summation factor. $a_n = 2a_{n-1} + 2^n$ with $a_0=0$ Divide by $2^n$

\[\frac {a\_n}{2^n} = \frac {a\_{n-1}}{2^{n-1}} +1\]

why 2^n? summation factor

For $a_n =x_na_{n-1}+….$

=> Divid by $x_n \cdot x_{n-1} \cdot x_{n-2}…..x_1$ $a =(1+1/n)a\_{n-1}+2$

Then summation factor: $\frac {n+1}{n} \cdot \frac {n}{n-1}....\frac {2}{1} = n+1$

therefore, $\frac {a\_n}{n+1} = \frac {a\_{n-1}}{n}+ \frac {2}{n+1}$

Types of recurrences

First Order

linear: $a_n = na_{n-1} - 1$

summation factor + telescoping => Divid by $x_n \cdot x_{n-1} \cdot x_{n-2}…..x_1$

nolinear: $a_n = \frac {1}{a_{n-1}}$

no-closed form solution

Simple Convergence

\[a\_n = \frac {1}{1+a\_{n-1}}, \quard n\>0 a\_0 = 1\]

Quadratic convergence and Newton’s method

Slow convergence

Second Order

Linear recurrences with constant coefficients)

\[a\_n = a_{n-1}+2a_{n-2}\]

1. GF

1.1 OGF for linear recurrences.

$a\_n = x_{1}a_{n-1}+x_{2}a_{n-2}+...x_{t}a_{n-t}$ $f(z) = g(z) \sum\_{0\leq n \< t} a\_n z^n (mod z^t)$ where $g(z) = 1- x_{1}z - x_{2}z^2….x_{t}z^t$ $f(z) = u_{0}(z)-u_{1}(z)...u\_{t}(z)$, dependens on initial values

$a_0,a_1,…a_{t-1}$ $a(z) = \frac {f(z)}{g(z)}$

for simple OGF, we can lookup the OGF table to get the a(n). However, sometimes, the a(z) is complexed expressed by z.

then we can expand generating functions or find functional equations on generating fucntions, to transform the generating function into the coeffient form of l(z)z^N Example Solving linear recurrence $a\_n = 5a_{n-1}-8a_{n-2}+4a\_{n-3},\quad for \ n \geq 3\ with \ a\_0=0,a\_1=1,a\_2=4$

Step 1: make recurence valid for all n

when $a_0=0$, it need no delta. when $a_1=1$, it need $\delta_{n1} = 1$, only works when n = 1 when $a_2=4$, it need $\delta_{n2} = -1$, only works when n = 2 So $a\_n = 5a_{n-1}-8a_{n-2}+4a_{n-3}+\delta_{n1}+\delta\_{n\_2}$

Step 2: multiple by $z^n$, and sum on n

since there will only a z for n = 1, and $-z^2$ for n = 2.
$A(z) = 5zA(z)-8z^2A(z)+4z^3A(z)+z-z^2$

Step 3: Solve A(z)

\[A(z) = \frac {z-z^2}{1-5z+8z^2-4z^3}\]

easy found that in the above formation of $\frac {f(z)}{g(z)}$

Step 4: Simplify or partial fractions.

$A(z) = \frac {z(1-z)}{(1-z)(1-2z)^2}$, sometimes, there will be partial fractions.

Step 5: Expand A(Z) to get a_n

\[a\_n = n2^{n-1}\]

some other application of generating functions: Probability Generating Functions to simplify the caculation for expectation and so on, Bivariate Generating Functions based couting and analysis of cost paramters.

2. Operate Method(Multiple roots $x^n$)

We have given a method for finding an exact solution for any linear recurrence. The process makes explicit the way in which the full solution is determined by the initial conditions. When the coefficients turn out to be zero and/or some roots have the same modulus, the result can be somewhat counterintuitive, though easily understood in this context

3. Analytic combinatorics

Symbolic methods, see more on Analytic combinatorics.

Step 1: Combinatatorial constructions
Step 2: Symblic transfer , Get GF equations.
Step 3: analytic transfer , Get coefficient asymptotic

variable coefficients: $a_n = na_{n-1}+ (n-1)a_{n-2} + 1$

summation factor $a\_n = na_{n – 1} + n(n – 1)a_{n – 2}$ => Divided by n!
Symbolic solution
GF and approximation methods

nolinear: $a_n = a_{n-1} a_{n-2} + \sqrt [2]{a_{n-2}}$

Higher Order

$a_n = f(a_{n-1},a_{n-2},a_{n-t})$

Full History

$a_n = n+a_{n-1}+a_{n-2}+…a_{1}$

Divide-and-Conquer

$a_n = n+a_{\lfloor{n/2}\rfloor}+a_{\lceil{n/2}\rceil}+n$

Classic examples:

* Binary search
* Mergesort
* Bather network
* Karatsuba multiplication
* Strassen matrix multiplication

Pattern:

Dividing into a parts of size about $N/\beta$
Solving recuresively
Combining solutions with extra cost $\Theta(N^{\gamma}(\log_N)^\delta)$

$a\_n = a_{n/\beta+O(1)}+ a_{n/\beta+O(1)}+....a\_{n/\beta+O(1)}+ \Theta(n^{\gamma}(\log\_N)^\delta$ with alpha terms for the sum, telescoping, ${\alpha}^{\log\_{\beta}^n} + \Theta (n^{\gamma}(\log\_N)^\delta)$ is given by $a\_n = \Theta (n^{\gamma}(\log\_N)^\delta) \quad when \quad \gamma \< log\_{\beta}^{\alpha}$ $a\_n = \Theta (n^{\gamma}(\log\_N)^{\delta+1}) \quad when \quad \gamma = log\_{\beta}^{\alpha}$ $a\_n = \Theta (n^{\log_{\beta}^{\alpha}} \quad when \quad \gamma \> log_{\beta}^{\alpha}$

More about Master Theorem, “A Master Theorem for Discrete Divide and Conquer Recurrences”, SODA 2011

Methods for solving recurrence

1. Change of Variable

2. Repertorie

Another path to exact solutions in some cases is the so-called repertoire method, where we use known functions to find a family of solutions similar to the one sought, which can be combined to give the answer. This method primarily applies to linear recurrences, involving the following steps:

Relax the recurrence by adding an extra functional term.
Substitute known functions into the recurrence to derive identities similar to the recurrence.
Take linear combinations of such identities to derive an equation identical to the recurrence.

The success of this method depends on being able to find a set of independent solutions, and on properly handling initial conditions. Intuition or knowledge about the form of the solution can be useful in determining the repertoire. The classic example of the use of this method is in the analysis of an equivalence algorithm by Knuth and Schönhage

3. Bootstrapping

Often we are able to guess the approximate value of the solution to a recurrence. Then, the recurrence itself can be used to place constraints on the estimate that can be used to give a more accurate estimate. Informally, this method involves the following steps:

Use the recurrence to calculate numerical values.
Guess the approximate form of the solution.
Substitute the approximate solution back into the recurrence.
Prove tighter bounds on the solution, based on the guessed solution and the substitution. For illustrative purposes, suppose that we apply this method to the Fibonacci recurrence: $a\_n = a_{n – 1} + a_{n – 2}$ for n > 1 with a0 = 0 and a1 = 1.

First, we note that an is increasing. Therefore, $ a_{n – 1} > a_{n – 2} and a_n > 2a_{n–2}$ Iterating this inequality implies that an > 2^{n/2}, so we know that an has at least an exponential rate of growth. On the other hand, $a_{n – 2} < a_{n –1}$ implies that $ a_n < 2a_{n – 1}, or (iterating) a_n < 2^n $. Thus we have proved upper and lower exponentially growing bounds on an and we can feel justified in “guessing” a solution of the form $a_n \sim c0{\alpha}^n$, where $\sqrt[2]{2} <\alpha < 2$ From the recurrence, we can conclude that a must satisfy $a^2 – a – 1 = 0$ Having determined the value a, we can bootstrap and go back to the recurrence and the initial values to find the appropriate coefficients.

4. Perturbation

Another path to an approximate solution to a recurrence is to solve a simpler related recurrence. This is a general approach to solving recurrences that consists of first studying simplified recurrences obtained by extracting what seems to be dominant parts, then solving the simplified recurrence, and finally comparing solutions of the original recurrence to those of the simplified recurrence. This technique is akin to a class of methods familiar in numerical analysis, perturbation methods. Informally, this method involves the following steps:

Modify the recurrence slightly to find a known recurrence.
Change variables to pull out the known bounds and transform into a recurrence on the (smaller) unknown part of the solution.
Bound the unknown “error” term.

Note2-AofA: A scientfic Approach

2015-09-19T14:05:00+00:00

A scientfic Approach

The scientfic method: O-notation, not at all useful for predicting performance

Scientific method calls for tilde-notion. Running time is ~aN^c, an effective path to predicting performance

Common error: Thinking that O-notaion is useful for predicting performance

Galactic algorithms:

R.J.Lipton: A galactic is one that will never be used. An effect would never be noticed in this galaxy. 75% SODA,95% STOC/FOCS are galctic

Steps

Analyze the algorithm by
- idenfiying anabstractr operation in the inner loop
- Develop a realistic model for input to the program.
- Analyze the frequeny of execution $C_n$ of op for input size N
Hypothsieze that the cost is ~ $aC_n$ where a is a constant
Validate the hypothesis by
- Developing generator for input according to model.
- Calculate a by running the program for large input.
- Run the program for larger input check the analysis
Validate the model by testing in application contexts.

Refine and repeat as necessary.

Tilde Notation.

Empirical:

* Run algorithm to solve real algorithm
* Measure running time and/or count operations Challenge: need good implementation

Mathematical:

* Develop mathematical model.
* Analyze algorithm within model Challenge: need good model. need to the math

Scientific:

* Run algorithm to solve real problem
* Check for agreement with model. Challenge: need all of the above.

Drawbacks:

Model may not be realistic.
- A challenge in any scientific discipline
- Advantage in CS: we can randomize to make the model apply.
Math may be too difficult
- A challenge in any scientific (c.f statistical physics)
- A “calculus” for AofA is the motivation for this course!
Experiments may be too difficult.
- Not compared to other scientific disciplines.
- Can’t implement? Why analyze?

symmetry:

$\sum_{1 \leq k\leq N} {C_{k-1}+C_{N-k}}$

both sums are $\sum_{1 \leq k \leq 2} C_{k-1}$

Quicksort compares: limiting distribution is not “normal” “Approximating the Limiting Quicksort Distributution”

Easy Method to Predict Hypothesis: Running time of Quicksort is ~aN in N.

Experiment.

* Run for input size N. Observe running time.
* Could solve for a.
* Predict time for 10N to increase by a factor ..

Validate-refine-analyze cycle

Note1-AofA: From AofA to AC

2015-09-19T11:33:00+00:00

This is an introduce lecture on Analysis of Algorithms. I will show the history and progress having made in the past, from Knuth’s scientific method, to Theory of Algorithm, and then to the current Analytic Combinatorial. We can get a general picture from this lecture.

#Introduction

Classic AofA Steps:

Develop recurrence relation
Derive GF equation
Extract coefficients
Asymptotics:Develop approximation

Context

Mathematics need

Recurrence
Genretating Function
Asymptotics
Trees
Permutations
Strings and Tries
Words and Maps

Ultimate Goal: Automatic Analysis

Analysis of Algorithm(1995) -> INRIA tech reports -> Analytic Combinatorics In principle, classical methods can provide

* full detials.
* full and accurate asymptotic estimates

In practice, it is ofter possible to

* generalize specialized derivations
* skip details and move directly to accurate asymptotics

Knuth: The Art of Computer Programming

To analyze an algorithm:

* Develop a good implementation and a realistic input mode
* Determine the caose and execution frequency of each operation.
* Calculate the total running time
* Run experiments to validate model and analysis  Beni: 

* Scientific foundation for AofA.
* Can predict performance and compare algorithms

Drawbacks:

* Model may be unrealistic.
* Excessive detail likely in the analysis

Theory of Algorithms (AHU 1970,CLRS)

To address Knuth drawbacks:

* Analyze worst-case cost [takes model out of the picture]
* Use O-notation for upper bound [takes detail out of analysis]
* Classify algorithms by these costs

Beni:

* Enable a new Age of Algorithm Design

Drawbacks:

* Analysis is often unsuitable for scientific studies (often overlooked)

Analytic combinatorics

can provide a basis for scitenfic studies.

* A calculus for developing models.
* Universal laws that encompass the detail in the analysis

Quantitative study of large combinatorial structures. Generating fuctions are the central object of study

AC Basic process:

Defin a combinatorial construction that precisely sepcifies the structure.
Use a symbolic transfer theorem to derive a GF equation.
Use an analytic transfer theorem to extract coefficient asymptotics

Combinatorial constructions:

* Algeraic formular built from natural combinatorial operators
* Operands are atoms or other combinatorial constructsions
* Two cases: atoms are unlabelled or labelled (all different)i ### Generating functions:
* controverial for some time. no particular meaning
symbolic meyhods: OGFs,EGFs,MGFs

Extract coefficients asymptotics

analytic transfer theorems based on GF as complex function.
Complex Asympotics: Singularity Analysis, Saddle Pointi [Asmptotiv
Counting,Moments of paramenters]
=> Random Structures: Multivarite Asymptotics, Singularity
Perturbation.[limit laws,Large Deviations]

combinatorial structures:

unlabelled universe:
    trees,strings,languages, compostitions.partitions, intergers.
labelled universe:
    permutations,cycles,words,mappings,urns,cayley trees

Universal laws

of sweeping generarlity are on hallmark of analytic combinatorics
Example: Context-free constructions
Grobner basis eleminations
Drmota-Lalley-wood theorem

Analytic combinatorics at the next level

Combinatorial parameters are handled with MGFS, often leadning to limit
laws.
Complicated singualrity strucuture leads to oscillatory behavior (like RS/PF
        formula in common).
GFs with no singularties require saddle-points asymptotics.
"If you can specify it, you can generate a random structure".
Analytic transfer understanding transformations form one combinatrorial
struture to another.
New types of implicit GS functional equations are arise. very strong list and growing

Applications of analytic combinatorics

* patterns in random strings
* polynomials over finite fields
* hashing
* data compression
* geometric search
* combinatorial chemistry
* arithmetic alogrithms
* planr maps and graphs
...

hadoop distcp from secure to insecure 0.20.2

2015-02-17T17:14:00+00:00

Background

When secure hadoop 0.20.2 want to access insecure hadoop 0.20.2, espeically for using distcp on hdfs, the job (distcp) will try to get token for both source and destination FileSystem. However, the insecure hadoop will return null token for disabled security option, which let secure hadoop client(DFSClient) throw NPE and job failed. Hence, we need to upgrade hadoop client code about getDelegationToken to work around this case.

Solution

Step 1

For null token received in DFSClient#getDelegationToken, not print it with stringifyToken

  
      // For insecure, result will be null, and throwing NPE by stringifyToken
      if(result != null){
        LOG.info("Created " + stringifyToken(result));
      }else{
        // for null token, handled by DistributedFileSystem for creating an dummy token.
        LOG.info("Created null token!");
      }

Step 2

For null token received in DistributedFileSystem#getDelegationToken, return a new dummy Token

   
   result = new Token(identifier,password,kind,service);
   
   </pre>
   

## Step 3 

Build new hadoop-core and deploy
         1. ant
      2. scp build/hadoop-core-0.20.2-cdh3u1.jar xxxx@xxxxx:/yourpath/
   

## Step 4

Set CLASSPATH using this new hadoop client jar.

   1. Only local support, let hadoop search HADOOP_CLASSPATH first for class loading.
          export HADOOP_CLASSPATH=/yourpath/hadoop-core-0.20.2-cdh3u1.jar
       export HADOOP_USER_CLASSPATH_FIRST=true
   

   2. For every node in secure cluster using this new jar.
   In most case, the job will get all required token, and set it into its Credential, which will make
   any other nodes in this cluster share this token.  So usually there is no need for every node upgrades this new client,
   which will also pollute the secure cluster. Please add the following code into the code of your hadoop job:
           job.getConfiguration().set("mapreduce.job.user.classpath.first", "true");
   


## Step 5
Just run your job as usually.

Google Analytics for Octopress

2014-04-26T21:52:00+00:00

Using Octopress as a blog framework is a funny experience, and how to display some blog stats such as pv is another interesting and challenge task for octopress post are only static html pages which means you cannot easily record the count by githubpages.

Google Analytics Service(GA) is a nature choice for web site analysis and there is already an Octopress plugin called jekyll-ga, which can sort blog posts by certain metrics of GA and another improved plugin called octopress-page-view. After trying these plugins, I jogged down some understandings and steps for how to using GA for Octopress as following:

1. What is Google Analytics

Do you know what people do when they visit your website or web app? Or how much the site contributes to your bottom line? Google Analytics keeps track and makes it easy for you to learn precisely what’s happening. Google Analytics shows you how to track different market segments and analyze conversion rates, and reveals advanced techniques such as marketing-campaign tracking, a valuable feature that most people overlook.

Accessing https://www.google.com/intl/zh/analytics/ and Using it is the best way to taste GA. An expert and hands-on book for GA can endow you a systematic view. The link on Amazon is the following. http://www.amazon.cn/mn/detailApp/ref=asc_df_0596158009956339/?asin=0596158009&tag=douban-23&creative=2384&creativeASIN=0596158009&linkCode=df0

2. How to track your site with Google Analytics

The above reference can offer your more details about GA. Here I just list some sensible cognition for understanding how Google Analytics works and how to set up GA service for your site.

Tracking ID

When you login to the GA console web page https://www.google.com/analytics/web/, you will asked to create a new project to tracking your site. First, a tracking ID and a segment of code will be generated. The tracking ID is the unique ID for your site, and it has the following funtions:

Notify the GA service when your site being accessed The generated code need to emebeded into your page template, so that the script will be called to notify GA with the tracking ID when someone is viewing your page. This just like the script is telling the GA,”Hello GA,someone in Japan is using the Chrome to view your client’s (TrackingID) page, please note down this.” So when GA receive this notice, it will record this viewing behavior in some database, and there comes a data design problem? How to organize the data?
TrackingID is the tableID You can imagine there is a data table used for these tracking record, and so a unique tracking ID is mapped into a unqiue table. Simplely speaking, the record will stored into a table called “TrackingID” and the data in this table will be analyzed and generated into a report including some graph displaying in the GA console. Some times you also need to fectch the GA data for other purposes using Google Analytics API but not only appretiate the generated report in the GA console.

Google Analytic API

With the generated trakcing ID, a table is created and tracking records are filled in. Then how to get these data is a coming issue the site analyzer will concern about. Of course, the greate google offer the Google Analytic API. More details about how to program with the API can be found here. https://developers.google.com/analytics/devguides/reporting/

More significantly, the lovely google offer a Google Analytic API service in the Google Developers Console: https://console.developers.google.com/project You can create a project and active the GA API service. The project is a proxy for the following funtions:

Authorize the clients who can send GA API request
The user should create a new client key in this GA API proxy and grant permissions for newly client. The generated client key contains a GA service account which should also be granted previleges in the GA console and a private certificate required by the client.
Forward the request to GA and return the response data to clients Now the client use the authorized GA service account and the private key to send request. The GA API service will forward the request from the client to the GA server to fectch ths data with tracking ID.

3. Get PageView Data For Octopress

Step 1. Installation

see more. https://github.com/developmentseed/jekyll-ga

Step 2. Set up a service account for the GA API

Turn on the Analytics API and accept the terms of service Go to API Access on the left sidebar menu, create a new oauth 2.0 client ID, give your project a name, and click next.
Select Application type: Service account, and click Create client ID
Note the private key’s password. It will probably be notasecret unless Google changes something. You’ll need to enter this value in your configuration settings.
Download the private key. Save this file because you can only download it once. Copy it to the root of your Jekyll repository. Safety tip: To protect this file, add its file name to your .gitignore file and to the exclude list in your _config.yml file
Note the Email address for the Service account. You’ll need this for your configuration settings and in the next step.
Log into Google Analytics and add the service account email address as a user of your Google Analytics profile: From a report page, Admin > User Management > Add permission for >

Step 3. Configure Page-View Plugin

see more. http://jhshi.me/2013/11/10/page-view-plugin-for-octopress/

#octopress-page-view page-view: service_account_email: #XXXXXX@developer.gserviceaccount.com key_file: privatekey.p12 #service account private key file key_file: privatekey.p12 #service account private key file key_secret: notasecret #service account private key's password profileID: #ga:XXXXXXXX start: 3 years ago #Beginning of report end: now #End of report metric: ga:pageviews #Metric code segment: #All visits optional filters: #optional

service_account_email, key_file, and key_secret come from the Google API console when you set up your service account.

profileID is the specific report profile from which you want to pull data. Find it by going to the report page in Google Analytics. Look at the URL. It will look something like https://www.google.com/analytics/web/?hl=en&pli=1#report/visitors-overview/###########p######/. The number after the p at the end of the URL is your profileID.

The start and end indicate the time range of data you want to query. They are parsed using Ruby’s Chronic gem, so you can include relative or absolute dates, such as now, yesterday, last month, 2 weeks ago. See Chronic’s documentation for more options.

The metric value is what you want to measure from your Google Analytics data. Usually this will be ga:pageviews or ga:visits, but it can be any metric available in Google Analytics. Specify only one. See the Google Analytics Query Explorer to experiment with different metrics. (Your dimension should always be ga:pagePath)

The segment and filters keys are optional parameters for your query. See the Google Analytics Query Explorer for a description of how to use them, or just leave them out.

The sort key can be true or false. If true, your posts will be sorted first by your Google Analytics metic, then chronologically as is the default. If false or not specified, your posts will sort as usual.

see more api usage. https://developers.google.com/analytics/devguides/reporting/

Step 4. Fetch Data and Generate Pages

As we all known, the github page only support the static pages, which means the page view count can only be generated at local and write into the generated html.

So rake generate will invoke this fechtin process in the pageview.rb written in ruby. You can read the code and add some debug code such as p results to print the request and the result in the console and to check whether the api request is right.

Here are some ways to check the rightness:

GA API console(https://console.developers.google.com/project) will show the request and the count of the error
Your debug code or some threw out excpetion info
watch your arguments and check in https://developers.google.com/analytics/devguides/reporting/

Octave on Mac

2013-12-14T19:21:00+00:00

For learning the online machine learning course on coursera, which use Octave for programming exercise, I need a Octave development enviroment on my MBP. But when I use brew install octave, I got a lot of errors, especially on Mac OS 10.9. After many times retry, eventually I sucesseed to install it, so I will depict my process on handle these problems occured in the tough installing process.


brew doctor
brew update
brew search octave
brew tap homebrew/science

0.Before using brew

Before using brew, I have tried to download the octave pkg on its official site. But this drag-and-drop version pkg is always crashed on my Mac OS 10.9. So maybe installed by compling source through homebrew is good idea. And then you will find you will get the following compiler not found error as following:


configure: error: no acceptable C compiler found in

For handling the error above, you should update your xcode and install the Command Line Tool. In the Mac OS 10.9, you need use xcode-select --install.

1.gfortran

By using brew install gfortran, you can succeed to installed it, but when brew install octave you will get the following error:


Undefined symbols for architecture x86_64:
"_append_history", referenced from:
_octave_append_history in liboctave_la-oct-rl-hist.o
(maybe you meant: _octave_append_history)
"_history_list", referenced from:
_octave_history_list in liboctave_la-oct-rl-hist.o
(maybe you meant: _octave_history_list)
"_read_history_range", referenced from:
_octave_read_history_range in liboctave_la-oct-rl-hist.o
(maybe you meant: _octave_read_history_range)
"_rl_basic_quote_characters", referenced from:
...
...

And I managed to solve this problem by installing a GCC compiler from the high performance computing website: http://hpc.sourceforge.net/

Now you can try brew install octave again. Then maybe you will get the following error again.


configure: error: A BLAS library was detected but found incompatible with
your Fortran 77 compiler settings

For this you should use another octave link as following:


brew install https://raw.github.com/Homebrew/homebrew-science/3c3fe3baaf926437f750f65456769c124d6be8e1/octave.rb --env=std

To here, every things seems ok. It will take several minutes to compile the whole octave.

3. Link octave to /usr/local/

Please pay attention that for gcc-4.8-bin.tar.gz the unziped destination fold is /usr/local. you need sudo, bu after that you will get every fold in /usr/local has a wrong pessmion, and when you use try brew install octave you will encouter the following error:


The formula built, but is not symlinked into /usr/local
Error: Permission denied - /usr/local/

So for permission issue, you can do sudo chown -R yourname:admin /usr/local/ and linke the octave again. To here, you already succeed a half and you can run octave in the the terminal now.

4. gnuplot

brew install gnuplot

Try run octave again, you will get this:


gnuplot> set terminal aqua enhanced title "Figure 1" size 560 420  font
"*,6"

For handling this error, you can type setenv("GNUTERM", "X11") before you use plot() in octave.(And of course you need install X11 first.)

Wish this passage can hlep you!

Mavericks CLT

2013-12-11T00:17:00+00:00

After the Apple Inc. make the Mavericks free, I upgrade my os of MBP into Mac OS X Mavericks.But I found that some basic header files such as “stdio.h” are gone and some tools can not be build normally.

So the problem is the Command Line Tools which has not be installed right, . For the early version of xcode. you can install the CLT by find it in the xcode->preferences->download. But for Mavericks, there is no more CLT in the download list. And the following is the right method to install.


xcode-select --install

Raspberry 无线网卡无法使用问题

2013-07-09T00:25:00+00:00

最近入手了Raspberry Pi玩下，一起也买了推荐的免驱的EDUP N8508GS, 买回来用了几天，实测确实免驱，不用自己折腾驱动了，刚开始使用的时候是将无线网卡接在HUB上的，使用没有什么问题，但是用了几天，突然wifi突然不能自动获取到IP了，设置静态的IP也不能访问网关，于是我试着重新切换到其他的系统，NOOBS确实是个入门尝鲜的好工具，可以很方便的安装各种raspberry pi的系统进行尝试，我依次尝试了raspbmc，raspbian,Pidora都不行。。。

后来发现，我将EDUP wifi网卡直接插在Raspberry的USB上，而不经过USB HUB,发现就可以了，目前还不清楚具体的是什么问题，看启动过程，插在hub上的时候，其实也是识别出了 wlan0网卡的，但是就是获取不到IP地址，而且手动配置也不行了，怀疑是电压或者hub本身的问题导致不能正常工作，同样有这样的问题，还有rapoo的键鼠套装。具体原因未知，遇到此类USB设备不好用的，可以试下直接连在Raspberry Pi的USB上试试，不经过USB HUB.

Jie Cao | Dialogue, NLP, ML

Rethinking On-policy Optimization for Query Augmentation

About the Work

Key Findings

Novel Contribution: OPQE

Authors

Links

Abstract

shortcut collections

Outline

Mac shortcut

Bash shortcut

Vim shortcut

IDEA shortcut

tmux shortcut

Note3-AofA: Recurrence

Telescoping a (linear first-order) recurrence

When the coeffients are 1

Elementary discrete sums

When the coffeients are not 1

why 2^n? summation factor

Types of recurrences

First Order

linear: $a_n = na_{n-1} - 1$

nolinear: $a_n = \frac {1}{a_{n-1}}$

Simple Convergence

Quadratic convergence and Newton’s method

Slow convergence

Second Order

Linear recurrences with constant coefficients)

1. GF

2. Operate Method(Multiple roots $x^n$)

3. Analytic combinatorics

variable coefficients: $a_n = na_{n-1}+ (n-1)a_{n-2} + 1$

nolinear: $a_n = a_{n-1} a_{n-2} + \sqrt [2]{a_{n-2}}$

Higher Order

Full History

Divide-and-Conquer

Classic examples:

Pattern:

Methods for solving recurrence

1. Change of Variable

2. Repertorie

3. Bootstrapping

4. Perturbation

Note2-AofA: A scientfic Approach

A scientfic Approach

Galactic algorithms:

Steps

Empirical:

Mathematical:

Scientific:

Drawbacks:

Note1-AofA: From AofA to AC

Context

Mathematics need

Ultimate Goal: Automatic Analysis

Knuth: The Art of Computer Programming

Theory of Algorithms (AHU 1970,CLRS)

Analytic combinatorics

Combinatorial constructions:

Extract coefficients asymptotics

combinatorial structures:

Universal laws

Analytic combinatorics at the next level

Applications of analytic combinatorics

hadoop distcp from secure to insecure 0.20.2

Background

Solution

Step 1

Step 2

Google Analytics for Octopress

1. What is Google Analytics

2. How to track your site with Google Analytics

Tracking ID

Google Analytic API

3. Get PageView Data For Octopress

Step 1. Installation

Step 2. Set up a service account for the GA API

Step 3. Configure Page-View Plugin