the daily grind - 3/3/11

Spent the day writing code. Partly successful, partly unsuccessful. Coz, I've implemented most of what I wanted to implement. There are a few more features to do, but they shouldn't take time.

But I'm stuck at generating graphs. Unfortunately, I can't seem to find a way around it.

The problem is as follows:

1) Since my procedure is stochastic, I have to run several simulations and then average out the numbers over the total simulations

2) The graph being drawn is a cummulative line. This implies, however, that the line can never go above the total number from which it is being drawn. Concretely, out of 30 agents, if the graph shows the number of agents self-adapting, it can never go above 30 in any round. The best that can happen, is that we hit 30 pretty quickly and then stay steady.

3) However, since we're summing up and averaging over 10 simulations, depending on how the algorithm played out, the number in any round becomes independent of the number from the previous round. That is, if the average number of successful agents in round 1 is 15, the number of successful agents in round 2 could be more than 15!

4) When you draw a cummulative line with data like that, the total number of agents goes above 30, which is clearly wrong!!

5) An option is to not draw a cummulative line. This would then show the number of agents that self-adapt in each round. This would solve the above problem, but takes away completely from the time element!

Hmmm...

the daily grind 24/2/11

Figured out two more papers to add to my lit review. I've already read them, so a once-through should be enough to summarise them and incorporate.

Spent more time on code.

Spent lots more time on interview prep. Found a really good book, which if I master should get me through at least some stages of the interview. Who knows, if I'll master it though. Fairly complex, but good stuff. Not too much time left. But then, even lesser time left for journal deadline.

the daily grind 21/2/11

A return to a typical PhD day :-(. Accomplished nothing.

After yesterday night's reading of papers, where I began to question my entire approach, this morning and afternoon were wasted, moping and worrying about it. Haven't done a spot of work. Tried to look at some code, but no point. Still worrying myself sick. Got to get out of this cycle. Have been down this road before, it's just depressing.

Might as well give up and go home. No point staring at a computer screen.

the daily grind 17/2/11

Although not much concrete work got done, I did do quite some amount of reading. Figured out two important papers that had slipped me by. Also found one on utility theory, that might directly impact my equations.

While washing dishes, I also came up with the idea of using the 'Byzantine Generals Problem' as a solution for gossip-based allocation of budget and QA. Basically, an agent should stop or start trading for the application, using the BG approach.

Got an email from Jenny, asking for dates for interview. Gave her dates in march, so that I can work on the journal paper. But, from say next week on, I'll have to devote evenings (at home) to studying the google-y topics of coding; also have to brush up on my java coding. Grrrr...

Tomorrow is likely going to be taken up by student evaluations. 5 hrs :-(. Will read a paper and a little bit more of that time mgmt thing.

the daily grind - 16/2/11

Spent most of today's working hours in meetings and student activities.

Was given a book called, "Getting Things Done" by a friend. Took a decision to spend some time reading it, just to see if I could make some changes to how I work. Will probably take it home and to bed with me.

Someone from Google is interested in talking to me, apparently. Have asked me for dates to interview. Will probably put off till march, so that I can prepare. Wonder if that's a bad idea, though it shouldn't be.

the daily grind - 15/2/11

Met with James Andrews today. He seemed interested in formulating my algorithm as a probabilistic game theoretic, well, formulation. But since he's really busy, any work on the proof will only happen by April. On a pessimistic note, should discount having anything regarding formal proof in my last RSMG report.

Spent most of the day preparing a covering letter and research statement for Harvard. A bit of a long-shot, really. But I think it's important to apply. Reach exceed grasp and all that.

Updated CV as a consequence, but not much else got done. Ah yes, read a bit about introduction to game theory :-)

the daily grind - 10/2/11

Started work on the journal article. Found the IEEE template

Emailed James about helping me with mathematical analysis. Fixed appointment for tuesday

Emailed Rami & Xin about reference letters for postd0c

Talked to Jenny from google

Tutor meeting for Msc. Java

Found reason for missing data points in plot. Fixed in one place. Need to fix in at least one more place.

the daily grind - 09/02/11

Worked on code to get the cumulative totalling of applications in the graphs. Not fully done yet, so not a fruitful day. Should've finished it, really. Been distracted.

But day isn't over yet. Will read up on gossip models and some associated literature, before going to bed.

Wonder how the HH manages to read so many papers/book chapters in one day.

7/2/11

Second successive day, lost to students and admin work!

Started off the morning with the tutorial that I had to give. Then worked on emails, for the plagiarism case that I detected.

The EPiCS meeting that I'm only 'guest attending' anyway, turned out to be two hours instead of one!

Gah! Not good at all. Shall try to finish a bit of coding before I sign off for the day.

3/2/2011

My day wasn't quite so productive. I spent it looking at the code and I think I know why things started working better. Remember I told you that there's buyers and sellers in my scheme of self-adaptation? Well, it turns out that in my code somewhere, I tried to be sophisticated and was making the sellers use a bi-modal distribution for creating their "asks", while the buyers were creating their "bids" using a normal distribution. Naturally, in the marketplace, the number of transactions were very limited! Also, to measure how much traders were deviating from the target, I was using an absolute number, which doesn't make sense. Deviation for each trader has to be measured relative to its own target. Once, I corrected that, the number of traders getting 'satisfaction' increased dramatically.

Hmmm...I started off thinking that I hadn't done much. But writing that down, makes me think that I did achieve something. It wasn't much, but at least there's some clarity and understanding of why things were working the way, they were.

I've still got to code up other composition types and verify that this scheme works for all of them, but I'm optimistic (hey! I'm a PhD student. I'm optimistic, by definition. Any sane person would've given up long ago ;-) ). Hopefully, I'll get at least something done tomorrow, though I don't have much hope, because I've got 3.5 hrs of student evaluation. Sigh.

1/2/2011

I don't understand it. I changed about three things and now the simulation works better than I expected!

If I spend a day figuring out why it works, this could be a breakthrough!

No further progress!

goals - 1/2/2011

I thought I'd follow Gemma's example and write down the set of short-term goals, I'd need to converge to a PhD.

So here goes:

1) Get code working for multiple composition types

2) Extend literature survey on self-adaptive services to qualify for a journal

#1. Write JouRNAL pAPer!!

3) Work on mathematical model for CDA-based adaptation

4) Write a qualitative review of CDA-based self-adaptation {contrast with other forms, perhaps}

5) Implement gossip algorithm for communication

#2. Write another paper - perhaps conference, perhaps journal

#3. Start writing thesis

6) thesis - Get literature review down

7) thesis - Description of CDA-based adaptation, design decisions

8) thesis - Evaluating this CDA-based adaptation

9) thesis - Introduction to thesis

31/1/2011

# Fixed the following bugs:

1) Sellers and buyers were using different distributions for generating asks/bids. Buyers were using normal distribution, while sellers were using a bimodal distribution. Now, both use normal distribution

2) Prevent sellers from bidding in auctions when there are no buyers present.

3) last_transaction is now set to None, at the beginning of a new trading round, rather than the end of the last trading round. This prevents buyers from getting a Null object while reformulating their bids. Behaviour seen since objects are not copied, objects are passed-by-value (python's consenting adults philosophy)

Not enough PROGRESS!! Some apps don't appear in the logs at all now. Very frustrating.

# Created a set of reading notes for MSc. student

# Created and delivered tutorial for tutees

# No response from James on algorithm. Perhaps I should wait for a couple of days and ping him.

the daily grind - 28/1/2011

Tried to figure out why buyers who weren't able to trade successfully, continued to be unsuccessful, even when successful buyers dropped out of the market.

Found bugs:

1) If no transaction happened, then modelling the buyer's next bid on the market's last transaction results in an empty bid

2) The sellers from previous successful transactions, drop out of the market, while the buyer's don't. This is the exact opposite of what we want.

Emailed James about help with describing the algorithm formally and doing some analysis (maybe probabilistic or game theoretic). Obviously, he won't read my paper immediately, but any critical response from him, over the next week, would be good.