ABOUT ME:

I am Hassan and since Nov. 2008, a Post-Doc with Prof. John Lloyd at CSL/RSISE, Australian National University, working on the project General Architecture for Artificially Intelligent Agents. I got my Doctorate from the Dept. of Computer Science at the University of Illinois at Urbana-Champaign in July/October 2008.
e-mail: hmahmud42 at gmail etc.

RESEARCH INTERESTS:

General Architecture for Artificially Intelligent Agents, Transfer Learning , Bayesian Machine Learning, Universal Prediction and application of Algorithmic Information Theory  to Machine Learning.

PAPERS:


Mahmud, M.M.H.: On Universal Transfer Learning. To appear in Theoretical Computer Science, 2009/2010.
(Longer version of the ALT paper).


Mahmud, M.M.H., Ray, S.: Transfer Learning using Kolmogorov complexity: Basic Theory and Empirical Evaluations.
To appear in, the Proceedings of the 20th Neural Information Processing Systems Conference, 2007.


Mahmud, M.M.H.: On universal transfer learning. In, the Proceedings of the 18th International Conference on Algorithmic Learning Theory, 2007.  Lecture Notes in Artificial Intelligence, LNAI 4754, pp. 135-149, 2007,  Springer, Berlin.

Mahmud, M.M.H., Ray, S.: Functional Similarity in Markov Environments
Workshop on Inductive Transfer, 18th Neural Information Processing Systems Conference, 2005.

Swarup, S, Mahmud, M.M.H., Lakkaraju, K,  and Ray, S: Cumulative Learning: Towards Designing Cognitive Architectures for Artificial Agents that Have a Lifetime Technical Report, University of Illinois at Urbana-Champaign, 2005. UIUCDCS-R-2005-2514.

Mahmud, M.M.H., Ray, S.: Using Functional Similarity to Transfer Information in Markov Environments
Unpublished, 2005

(Theory for the above paper)
Mahmud, M.M.H., Ray, S.: A Novel Forward Model for Markov Environments
Unpublished, 2005.

Ph.D. Thesis

Mahmud., M. M. H. Universal Transfer Learning. Ph.D. Thesis, University of Illinois at Urbana-Champaign, 2008.

In addition to the material in the NIPS paper and ALT paper, the dissertation contains full development of parallel transfer, competitive optimality of universal priors, Kolmogorov complexity of functions, and many, many more experiments.

M.S. Thesis

Mahmud., M. M. Hassan.  Explanation Based Policy Adaptation. Master's Thesis, University of Illinois at Urbana-Champaign, 2002.

We derived a method that adapts a policy learned in an ideal setting using prior knowledge, so that it works in the actual setting. So this way we require fewer actual examples to learn the policy. We applied our method in a simulated Air Hockey robot problem (simulated using equations derived for an actual robot), which is an example of a complex non-linear dynamic control problem.

Please e-mail me if you have any problems downloading, or have comments, questions etc.

RESEARCH INTERESTS IN MORE DETAIL:

General Architecture for Artificially Intelligent Agents

We have ideal but non-computable architectures for AI agents (such as AIXI) and practical but small scale ones such as the MDP based planners popular in machine learning community. To develop an architecture for agents that can handle multiple complex tasks, in parallel or in sequence (such as robotic helpers in the context of some class of applications, or majordomo software agents to manage a household) and that can also be used easily with a minimum of handcrafting, we need to strike a balance between these two extremes. This project is about trying to figure out how to do this.

Transfer Learning and AIT.

Formally solving various problems in Transfer Learning using AIT, and developing powerful practical algorithms by approximating the formal solutions (see papers). Application of these ideas to Transfer Learning in Universally Artificially Intelligent agents.


More sophisticated approximation to the universally optimal transfer methods

The transfer methods we have developed are computable only in the limit. So we need to approximate them in practice, and the approximations we have developed so far, while effective, are not as sophisticated as we would like them to be. Furthermore, we have only applied our method to standard UCI test datasets. We would like to rectify the above two issues and construct more sophisticated approximation to our transfer method, and apply them to novel and challenging real life applications. It seems the most fruitful path to pursue from both a theoretical and practical perspective is to develop approximations tailored to particular class of domains: e.g. bioinformatics, medical diagonistics, vision, cheminformatics, analysis of financial markets/economics, robotics - essentially approximation of our method tailored for any application domain of machine learning.


Transfer Learning in Bayesian Reinforcement Learning Agents.

If we consider Factored Bayesian MDPs then the transfer method developed in our NIPS paper can be easily extended to reinforcement learning. In essence, the model for the MDP is represented by a decision tree, or a belief net, and since we use decision trees in our NIPS paper, we can use the same method to transfer in this setting too. That is, we can learn the model of a novel MDP quicker using the models of previously learned tasks.


Resource Bounded Bayesian Learning

The goal of this area is to investigate the construction of learning/prediction algorithms that are resource boundedly Universal - i.e. predictors that are guaranteed to not do much worse than any other predictor using the same amount of resources (time and memory). The main attraction of this approach is that successful research would provide a principled way to automatize machine learning - the only parameter the user need specify is the available resources and he/she is guaranteed performance quite close to the optimal.

Recent exciting work in this direction include:

Meron E. and Feder M. Finite-Memory Universal Prediction Of Individual Sequences, IEEE Transactions on Information Theory 50(7): 1506-1523 (2004)

Ingber A. and FederM. Non-Asymptotic Design Of Finite State Universal Predictors for Individual Sequences, DCC 2006: 3-12

J. Ziv and N. Merhav, ``On context-tree prediction of individual sequences,'' IEEE Trans. Inform. Theory, vol. 53, no. 5, pp. 1860-1866, May 2007.


We would like to consider Bayesian resource bounded learning in a much more general setting than in the above papers. Here the set of probability measures are computable probability measures respecting a given resource constraint - i.e. probability measures that are computed by programs using time and memory T,M. Unfortunately, in this case, it is easy to establish that we cannot use the techniques used in the above papers. However, one approach that holds promise is using MCMC over program space. MCMC Convergence rates established for structured/combinatorial state spaces seem to indicate that we should be able to derive convergence rates in our case as well. Of course this is just an initial idea, and a lot of work needs to be done to establish this algorithm as being resource boundedly universal.