Changing the Blog's Screen Width

http://bloggersentral.blogspot.com/2009/06/how-to-change-blog-width.html

Learning in Games - Some Notes

Strategic Learning and its limits - Peyton Young

One school points optimistically to particular classes of games that can in fact be learned by simple updating procedures, such as learning minimax equilibria by fictitious play (see Chapter 6). This school also points to the fact that any game can be learned by Bayesian methods provided that the priors are sufficiently aligned to begin with (Kalai and Lehrer, 1993).
The other school views the situation more pessimistically. Its adherents point out that simple learning procedures tend to work only in special cases, such as zero-sum games, potential games, and so forth. Moreover, they are unimpressed by the fact that Bayesian methods lead to learning if the priors are sufficiently aligned to begin with. First, Bayesian reasoning makes extreme demands on the computational capacities of the players, and hence is not a plausible model of decision making by human beings. Secondly, even if one is willing to overlook the empirical implausibility of this approach, the amount of coordination required to learn equilibrium effectively assumes some notion of pre-equilibrium in the prior beliefs. Hence it does not solve the learning problem in a robust sense. (These matters are treated in Chapter 7.)
The purpose of these lectures is to survey results on both sides of the question. On balance, I shall give a little more weight to the positive side, though not in a way that will please all of the positivists. I begin by examining simple forms of learning behavior that depend only on realized payoffs and simple statistical summaries of historical data. These include forms of reinforcement and no regret learning, which turn out to be closely connected. Some of these rules converge in long-run average behavior to correlated forms of equilibrium, including standard correlated equilibrium and variants thereof (see Chapters 2–5). An important property of these learning rules is that they are robust, that is, they require no prior knowledge of the opponents' payoffs or strategies. Significantly, however, they do not yield Nash equilibrium behavior except under special circumstances.
In Chapter 6 we study fictitious play and its variants. These can be viewed as primitive forms of Bayesian learning in which players use simple statistical models based on past evidence and choose best (or almost best) replies given the evidence. Smoothed versions of fictitious play have long-run average behavior that is similar to that of no regret learning, but they do not converge to Nash equilibrium except in special cases.
We then examine the case in which players are very sophisticated in order to see how much can be gained from a high rationality perspective (Chapter 7). The major positive result here is due to Kalai and Lehrer (1993), who show that Bayesian learning leads to Nash equilibrium.
Our concern here is with the learnability of Nash equilibrium in games with finite action spaces. If one allows infinite action spaces, there exist games in which one cannot even determine in finite time whether a Nash equilibrium exists, let alone learn it by a decentralized process (Rabin, 1957; Prasad, 1991, 1997).

install acroread in fedora

wget http://ardownload.adobe.com/pub/adobe/reader/unix/8.x/8.1.2/enu/AdobeReader_enu-8.1.2-1.i486.rpm
yum -y localinstall --nogpgcheck AdobeReader_enu-8.1.2-1.i486.rpm
/bin/rm AdobeReader_enu-8.1.2-1.i486.rpm

Debian commands

To have alternate versions of the same program like matlab, java, etc
man 8 update-alternatives
/usr/sbin/update-alternatives --list matlab


To install a new font (true type) put the font in /usr/share/fonts/truetype/ (in debian) and run
fc-cache -fv
then restart X ( /etc/init.d/gdm restart if using GNOME display manager)


To sort folders in current directory by size
du --max-depth=1 ./ | sort -n -r

Some good technical links

For fixing NTFS not mounting in Linux (Debian) check this link :http://ubuntuforums.org/showthread.php?t=656025&page=2

how to install any package in latex

create a texhash directory in home folder. download the package say "latex-beamer.ter.gz" to a folder inside the texmf/ tree.say "texmf/latex/beamer/". now all the contents of the package should be put in this directory.
Then run "texhash" for latex to find these files.

if you have any dtx file in the package like for eg. "empheq.dtx" then you need to run "tex empheq.dtx" to create the runtime environment files.

wireless cards supported on linux

REFERRED FROM : http://sushant354.blogspot.com/2007/02/wireless-card-supported-for-linux.html

Linux limited support for wireless devices makes one often switch to windows. I often wanted wireless cards that work on linux without any effort. If you want a wireless card to work without any problem, buy a Prism chipset supported on linux exclusively at
http://prism54.org. The card that I have is a Netgear WG511 (Made in Taiwan and not China).

Yesterday I also got a Ralink chipset exclusively supported at http://rt2x00.serialmonkey.com/wiki/index.php?title=Main_Page for just $16.

My card is Gigabyte WMKG and it works flawlessly on linux. You can find other cards with Ralink chipset at http://ralink.rapla.net/

You may also find a huge bunch of Atheros chipset being supported at http://www.madwifi.org.