Abstract

Text summarization is the process of distilling the most important information from a source to produce an abridged version for a particular user or task. That is restating of the main ideas of the text in as few words as possible in a pertinent format. Automated text summarization has drawn a lot of interest in the natural language processing and information retrieval communities in the recent years due to the rapid growth of the World Wide Web and electronic information services. Information is becoming available on-line at an incredible rate but no one has time to read everything. Hence, the technology of automatic text summarization is becoming indispensable for dealing with this problem. With the growing need for good summarization systems, summarization evaluation plays an important role in their creation. Evaluation has become one of the ongoing issue in this field of automatic text summarization. Human judgement often has wide variance on what is considered a ”good” summary, which means that making the evaluation process automatic is particularly difficult. Manual evaluation like ’Likert Scale rating’, ’Pyramid Evaluation’ can be used, but this is both time and labor intensive as it requires humans to read not only the summaries but also the source documents. Hence, there is an urgent need for developing automatic evaluation metrics to fasten the process of evaluating summarization systems which can result in better summarizing algorithms. Summaries can be evaluated based on two properties : content and form. The evaluation of content is called ’Content Evaluations’ and that of form is called ’Readability Evaluation’. There are various manual evaluation metrics for content evaluation including ’Likert Scale rating’, ’Pyramid Evaluation’ and ’Content Responsiveness’. For Readability evaluations a ’Likert scale rating’ is used by human experts to rank summaries. In the context of automated metrics for evaluation of summary quality, content evaluations have been presented through the last decade and continue to evolve. However, a careful examination of readability aspects of summary quality has not been as exhaustive. In this thesis, we dissect readability into 5 parameters and automate the evaluation of readability based on some of these parameters. Readability of a summary is usually graded manually on five aspects: grammaticality, coherence, focus, referential clarity and nonredundancy. We explore alternative evaluation metrics for grammatvviicality, focus and coherence. By defining grammaticality as a degree of acceptance, we use language models built over shallow features such as parts-of-speech (POS) and chunk tags to estimate the acceptability of a sentence to be a part of the summary. To capture the coherence in a summary, we use lexical chains and compute semantic relatedness of the sentences within the summary. We claim that the focus of a summary is a characteristic dependent on the cohesiveness of a summary, due to which, any metric for coherence should be a good indicator of focus. Our approaches to capture grammaticality, focus and coherence are able to strongly correlate with manual ratings. Overall, in context of query focused multi-document summaries, we observe that none of the five aspects of readability are independent of each other, and hence by addressing the individual criteria of evaluation, we aim to achieve automated appreciation of readability of summaries.

Abstract

Data Mining is a technology that blends traditional data analysis methods with sophisticated algorithms for processing large volumes of data. It is the process of automatically discovering useful information in large data repositories. Data Mining includes extensively researched techniques like clustering, outlier mining, frequent itemset mining, classification and regression. In this thesis, we consider the problem of extracting the special properties of any given record in a relational dataset. We are interested in determining what makes a given record special or different from the majority of the records in a dataset. In the real world, records typically represent objects or people and it is often worthwhile to know what special properties are present in each object or person, so that we can make the best use of them. This problem has not been considered earlier in the research literature and led to the introduction of Speciality Mining. There are several scenarios where this Speciality Mining can be useful: A shopper can select from a range of products by analysing the special properties of each product in comparison with the other products. For example, there are many varieties of mobile phones in the market and it would be convenient for the customer to know how each mobile phone differs from the remaining group. “Speciality Mining” is based on the intuition that a record is special if it contains a property that is present in very few records. Intuitively, we say that a property is special if it is present in only a few objects. However, we argue that if similar properties are present in a large number of objects, then even such a property should not be considered special. For example, in a database of 20 people if only 2-3 persons know English, they are special with respect to the language attribute. However, if for every language in the database, there are only 2-3 people who know that language, then it means that all people are special with respect to the language attribute. This is absurd – if everyone is special, no one is truly special. To ensure that the above definition of speciality captures this intuition in an elegant fashion, we use ideas from information theory, cluster analysis and frequent itemsets. We also provide the speciality of records in a convenient, easily-interpretable manner. An illustrated example is provided on how users can interpret the results. Experiments on real and synthetic data sets reveal interesting properties of data records that cannot be mined using traditional approaches. Keywords: Cluster analysis, Frequent itemsets, Outlier, Surprisal.

Abstract

The exponential growth of world wide web has transformed it into an ocean of knowledge in which highly diverse information is linked in a extremely complex and arbitrary manner. It has become the repository for all the information needs of the users who search for answers to their diverse needs. Having been entrusted with the responsibility of serving the information needs of this ever- increasing pool of internet users, search engines, in order to sustain this highly competitive and huge market, are forced to mine this vast pool of information to serve information that is relevant, accurate and veritable. Typically, a search engine in these days, returns hundreds of results in response to a user query. The user is left with the painful task of mining these results to find the information needed. But sometimes, these results may not contain the information needed by the user and hence the user has to rephrase the query and restart the search session. The search engine which minimizes this process and makes it more user-friendly is bound to capture more market. Traditional search engines used to return generalized results for particular queries i.e., they used to rank the results based on query irrespective of the preferences of the searcher. But such generalized ranking of results may not satisfy all the searchers since each searcher is unique and has different information need(sometimes even in case of similar queries), different tastes and preferences. Hence, of late, search engines have started concentrating on fetching personalized results according to the taste of each individual user. For example, for a query “Sourav ganguly”, a user who likes both “Sourav Ganguly” and “Shahrukh khan” may be more interested in news about “IPL”, which has both the subjects involved in it, when compared to a user who does not like “Shahrukh khan”. Similarly, for a query “Apple”, a user who is a computer professional may like to know news about “Apple products” whereas a botanist or a dietician may want to know about “Apple” fruit. Thus search engines started focussing on knowing the tastes and preferences of the users and retrieve results that they might find more relevant. Personalization of search can be done by knowing information about the user like his topics of interests, his background knowledge on different topics and his browsing model. This user information can be known by posing a generalized set of questions to the user through which we can learn the information we need about the searcher which is called explicit feedback. But again, such generalized set of questions shall limit the extent to which we can learn the searcher and also such feedback is expensive to collect due to the amount of effort of the user involved. Hence researchers have sorted out a new type of feedback of late. Here, we collect the user interactions with the search engine without any effort from the user. Such feedback is called Implicit relevance feedback. Implicit relevance feedback has received wide attention recently as a means to capture the search context and improving search results. However, implicit feedback is usually not available to public or even research communities at large for reasons like being a potential threat to privacy of web users. Hence, researchers in personalization of search are compelled to generate their own datasets in lab environments. These datasets are generally domain-centric, uniform and restricted to a small set of users. But if we examine the feedback data from a real world commercial search engines, it may range over a number of domains, varying in subject and content and may involve millions of users. Thus, personalization algorithms trained on lab generated dataset may not satisfy a real world search user. Hence, it is difficult to experiment and evaluate web search related research, especially web search personalization algorithms. Given these problems, we are motivated towards creating a synthetic user relevance feedback data, based on insights from query log analysis. We call this Simulated feedback. We believe that simulated feedback can be immensely beneficial to web search engine and personalization research communities by greatly reducing efforts involved in collecting user feedback. The benefits from ”Simulated feedback” is that - It is easy to obtain and also the process of obtaining the feedback data is repeatable, customizable and does not need the interactions with the user. In this thesis, we describe a simple yet effective approach using Prognostic Search process for creating simulated feedback. In this approach, we create virtual users from the user profile data of real world users and simulate search sessions of these users. We propose two models to represent the user profile of a virtual user viz., “Skewed Mean Model” and “Probabilistic Bayesian Model”. We achieved high accuracy levels with both the models. But the

Abstract

Large scale pattern classification systems are necessary in many real life problems like object recognition, bio-informatics, character recognition, biometrics and data-mining. This thesis focuses on pattern classification issues associated with character recognition, with special emphasis on Malayalam. We propose an architecture for the character classification, and proves the utility of the the proposed method by validating on a large dataset. The challenges we address in this work includes: (i) Classification in presence of large number of classes (ii) Efficient implementation of effective large scale classification (iii) Simultaneous performance analysis and learning in large data sets (of Millions of examples). Throughout this work, we use examples of characters (or symbols) extracted from real- life Malayalam document images. Developing annotated data set at the symbol level from a coarse (say word-level) annotated data is addressed first with the help of a dynamic program- ming based algorithm. Algorithm is then generalized to handle the popular degradations in the form of cuts, merges and other artifacts. As a byproduct, this algorithms allows to quantitatively estimate the quality of the books, documents and words. The dynamic programming based algorithm aligns the text (in UNICODE) with images (in Pixels). This helps in developing a large data set which could help in conducting large scale character classification experiments. We then conduct an empirical study of classifiers and feature combination to explore their suitability to the problem of character classification. The scope of this study include (a) applicability of a spectrum of popular classifiers and features (b)scalability of classifiers with the increase in number of classes (c) sensitivity of features to degradation (d) generalization across fonts and (e) applicability across scripts. It may be noted that all these aspects are important to solve practical character classification problems. Our empirical studies provide convincing evidences to support the utility of SVM (multiple pair-wise) classifiers for solving the problem. However, a direct use of multiple SVM classifiers has certain disadvantages: (i) since there are nC2 pairwise classifiers, storage and computational complexity of the final classifier becomes high for many practical applications. (ii) they directly provide a class label and fail to provide an estimate of the posterior probability. We address these issues by efficiently designing a Decision Directed Acyclic Graph (DDAG) classifier and using the appropriate feature space. We also propose efficient methods to minimize the storage complexity of support vectors for the classification purpose. We also extend our algebraic simplification method for simplifying hierarchical classifier solutions.We use SVM pair-wise classifiers with DDAG architecture for classification. We use linear kernel for SVM, considering the fact that most of the classes in a large class problem are linearly separable. We carried out our classification experiments on a huge data set, with more than 200 classes and 50 million examples, collected from 12 scanned Malayalam books. Based on the number of cuts, merges detected, the quality definitions are imposed on the document image pages. The experiments are conducted on pages with varying quality. We could achieve a reasonably high accuracy on all the data considered. We do an extensive evaluation of the performance on this data set which is more than 2000 pages. In presence of large and diverse collection of examples, it becomes important to continu- ously learn and adapt. Such an approach could be more significant while recognizing books. We extend our classifier system to continuously improve the performance by providing feed- back and retraining the classifier. We also discuss the limitations of the current work and scope for future work.

Abstract

The signi¯cant growth in computational power of modern Graphics Processing Units (GPUs) cou- pled with the advent of general purpose programming environments like NVIDIA's CUDA, has seen GPUs emerging as a very popular parallel computing platform. A lot of work has gone into successfully implementing a large number of algorithms on the GPU. While there is a general consensus on what types of algorithms can successfully leverage the computing power of the GPU's, there is no way to ascertain how much faster the algorithm will be as compared to a sequential version or how it will compare with an alternative algorithm for the same task. In other words this is the analytical equivalent of stating that The only way to ¯nd out how fast an algorithm is, is to implement it and measure it's running time". This clearly demonstrates a huge gap in our understanding of GPGPU computing. A performance prediction model for the CUDA GPGPU platform is presented in this thesis. This model encompasses the various facets of the GPU architecture like scheduling, memory hierarchy, and pipelining among others. We also perform experiments that demonstrate the e®ects of various memory access strategies. The proposed model can be used to analyze pseudo code for a CUDA kernel to obtain a performance estimate, in a way that is similar to performing asymptotic analysis. To validate the model the analysis of three fundamental Parallel algorithms is performed: Matrix multiplication, List ranking, and Histogram generation. To further illustrate the power of our model, two symmetric key encryption algorithms Blow¯sh, The Advanced Encryption Standard (AES) and their secure versions are implemented and analyzed. These encryption algorithms prove to be particularly close to most real world algorithms as they contain a mix of heavy computation, irregular memory access and the movement of large amounts of data.

Abstract

Words can have more than one distinct meaning and many words can be interpreted in multiple ways depending on the context in which they occur. This phenomena poses challenges to Natural Language Processing systems. State-of-the-art methods to resolve word ambiguity make use of manually annotated data. However, obtaining such data is costly for certain languages and domains. In this thesis we have developed word sense disambiguation (WSD) methods for languages and domains where no annotated data is available. We have proposed unsupervised corpus based methods for Semantic Category Labeling, a task very similar to assigning coarse grained WSD and hence relatively easier than traditional WSD. Methods that rely on lexical knowledge base (WordNet) are also evaluated. Furthermore, we have developed methods for domain-specific WSD and evaluated our performance on “domain specific WSD of all words” task which is a part of ACL SemEval- 2010. The results reveal the importance of domain information in domain-specific WSD. Very little work has been reported on integrating various knowledge sources for WSD. Current methods for WSD take advantage of only a few knowledge sources and do not use them collectively. We propose a novel framework which can model information from various knowledge sources into constraints and collectively use them for disambiguation. Our initial experimental results are competitive with that of state-of-the-art methods.

Abstract

Automatic speech summarization is the task of generating a concise summary of a speech signal using a digital computer. The existing speech summarization systems rely on automatic speech recognition (ASR) transcripts and gold standard human summaries to generate summaries of speech signals. The limitations with these approaches are, ASR errors make summaries less usable by humans, also ASR systems are not available for all languages, especially for less resource languages and it takes considerable resources and effort in building one. Gold standard human summaries are not available for all speech signals and building them is tedious and time consuming task. In this work, we propose two techniques for summarization: 1) Exploiting anchor speaker role in broadcast news (BN) show to construct summaries, 2) A generalized ranking of speech segments based on prominence values of syllables in them. By analyzing manual summaries of news shows, it was found that anchor speaker segments are mostly picked in manual summaries. Therefore it is desirable for automatic summaries to exhibit this characteristic. We proposed two techniques to perform anchor speaker tracking, based on auto associative neural network model and Bayesian information criterion method. Audio summaries are generated for desired summary length by concatenating anchor speaker segments based on their positional features. These summaries are evaluated with the help of ROUGE, an automatic text summarization evaluation pack age by transcribing the audio summary into text. ROUGE-N metric measures the N-gram overlap between human reference summaries and the automatic summary. The f-measure scores of the proposed system for ROUGE-1 and ROUGE-2 metrics are 0.561 and 0.392 respectively. These scores showed that the system is capable of generating summaries that are as good as supervised speech summarization system trained using gold standard human summaries which achieved 0.553 and 0.382 for ROUGE-1 and ROUGE-2 metrics respectively. Also, we performed a task based evaluation where, humans were asked to listen to the summary and answer questions regarding the contents of a news show. The percentage of questions answered by the humans was 71 % for the proposed system which is better than 60.2 % of the supervised speech summarization system. The coherence of the summaries was also evaluated by asking the users to rate the summaries on a scale of 1-5 where 1 corresponds to very bad and 5 corresponds to very good. The mean opinion scores (MOS) of these ratings for the proposed method and the supervised speech summarization system are 4.05 and 3.2 respectively. The task based evaluation of these summaries by humans showed that, they prefer the summaries generated by the proposed techniques over the summaries generated by standard speech summarization methods. In other part of the work, a technique to rank segments in a speech signal using prosodic features that indicate importance is proposed. When humans convey message through speech, they attract listeners’ attention to information bearing parts of speech through variations in pitch, amplitude, duration and stress. Speakers make some words prominent and reduce other words. The proposed method computes syllable level prominence values as a function of syllable nucleus duration, sub-band energy (300-2200 Hz), and pitch variation and these values are used to obtain a segment level score, which is used for ranking the segment for summarization. It is shown that this type of scoring captures the prosodic information relevant to summarization in an unsupervised framework. We have also proposed a method to combine lexical and positional features with the prominence based scoring when text transcripts of speech signals are available. The proposed prominence based scoring captures complimentary information to lexical features derived from text transcripts of speech signals. The combination of these features perform better than the individual features. The proposed method was evaluated on two types of speech data; read style news speech and spontaneous telephone conversations. The proposed system based on prominence scoring achieved ROUGE-1 and ROUGE-2 f-measure scores of 0.508, 0.341 on read style news speech and 0.666, 0.464 on spontaneous conversations respectively. In read style speech the basic unit of extraction was obtained based on pause based segmentation which does not give semantically meaningful segments, where as in spontaneous telephone conversations we have considered speaker turns which are semantically meaningful units as basic unit of extraction.

Abstract

Conventional CMOS logic circuit design approaches rely on charging the output capacitive nodes to the power supply Vdd or discharging it to ground. This design approach has been widely used. For Low power applications, although there are many techniques both at circuit level and system level, a very fundamental source of energy dissipation is the discharge of the capacitor to ground. Every time a capacitor is discharged to ground, an amount of energy = 12 CV 2 stored in the capacitor is lost. This loss of energy can be prevented if instead of discharging the capacitor to ground, the charge or energy stored can be recycled. Another way of reducing the power consumption is to design the circuit in such a way that the charging of the capacitive node takes place very slowly. It has been observed that slowly charging the capacitor requires lesser energy than abrupt charging. Adiabatic circuits use the above two methods viz. slow charging and discharging, and charge recycling to minimize the power consumed. Several Adiabatic designs have been proposed in literature. Most of them achieve significant power savings in comparison to conventional circuits. Drawbacks of these circuits include complex circuit design for achieving simple operations, requirement of multiple clocks and/or requirement of complimentary input signals for controlling the charging/discharging process. The Current work is based on an existing adiabatic logic style known as GFCAL (Glitch Free and Cascadable Adiabatic Logic )which is simple and does not require complimentary signals or complex clocking. The GFCAL has a drawback that it uses junction diodes for controlling the charging and discharging of output nodal capacitance. Junction diodes are difficult to fabricate in a CMOS process and they are responsible for power dissipation due to the cut-in voltage drop across them. This work improves the circuit by replacing the diodes with transistors that control the charging and discharging process. The efficiency has gone up from 50% for GFCAL to about 60% for the modified circuit in comparison to conventional CMOS logic circuits. The modified design is found to work satisfactorily, is cascadable and simple block level implementations like adders have been demonstrated and verified to work with lower power consumption than conventional CMOS circuits.

Abstract

The technology evolution and complexity of new circuit applications involve emerging reliability problems and even more, sensitivity of integrated circuits (ICs) to electrostatic discharge (ESD) and Overvoltage stress-induced damage. Regardless of the aggressive evolution in downscaling and subsequent improvement in applications’ performance, ICs still should comply with minimum standards of failure robustness in order to be commercially viable. Although the topic of IC Protection from ESD and Over voltage induced damages has received attention industry-wide, the design of robust protection structures and circuits remains challenging because the failure mechanisms continue to become more acute and design windows less flexible. The sensitivity of smaller devices, along with a limited understanding of the ESD phenomena and the resulting empirical approach to solving the problem have yielded time consuming, costly and unpredictable design procedures. As turnaround design cycles in new technologies continue to decrease, the traditional trial-and-error design strategy is no longer acceptable, and better analysis capabilities and a systematic design approach are essential to accomplish the increasingly difficult task of adequate protection against IC failures. This dissertation presents comprehensive design methodology for implementing on-chip ESD protection for LNA and a circuit design technique to overcome Reliability issues in Mixed Voltage IOs. Firstly the various failure mechanisms in Integrated circuits is revised. In the next chapter, the ESD topic in the semiconductor industry is described, as well as ESD standards and commonly used schemes to provide ESD protection in ICs. The general ESD protection approach is illustrated and discussed using different types of protection components and the concept of the ESD design window. The problem associated with capacitive loading of ESD protection device in LNAs is addressed next. A better methodology for implementing greater level of ESD protection by making use of ON chip inductors is proposed. The passives used in the design of RF Receiver blocks also influence the ESD robustness of LNA. Inductors and capacitors in RF circuits can be implemented either ON chip or OFF chip. In earlier technologies where the passives used to be OFF chip the ESD immunity of the LNA was totally determined by the robustness of ESD protection device. However in modern RF ICs the focus is on full integration to reduce system cost and area. So the passives appear ON chip. In RF ESD Protection it is common practice where capacitors and inductors in the circuit used for matching and frequency selection, are used to tune out ESD frequencies. In a Cascode LNA with inductive degeneration where ON chip inductors are used they can be used for ESD protection. Two cases are considered. In the first case a common design practice where the ESD device is positioned at the gate of LNA is considered as shown in Figure 1.1 and it is shown that the level of ESD protection is very much limited as the frequency of operation increases. In the second case we shift the ESD device away from the gate of LNA and make use of ON chip inductors to enhance the level of ESD protection. This method also has been shown to that the advantage that the capacitive effects are reduced as we move away from the gate of LNA. Consider the topology in Figure 1.1 (a).As the frequency increases the capacitive reactance decreases and the ESD device begin to shunt more signal power to the ground. With this topology around 5GHz frequency of operation the achievable level of ESD protection may be reduced to 500V HBM which leads to early failure of the circuit.The proposed methodology for providing ESD protection for LNA using ON chip inductors is shown in Figure1.1 (b).By shifting the ESD protection device away from the gate of LNA and towards antenna the capacitive effects are reduced which means that the degradation of power gain and noise figure is less for given ESD device size. Also the required matching of 50Ω for maximum power transfer can achieved by varying both gate inductor Lg and source inductor Ls providing higher degree of freedom. Also the presence of gate inductor suppress steep current transients. With the proposed methodology upto 4KV HBM ESD protection is shown to be possible without much degrading the LNA performance parameters where as with the existing approach maximum protection of 2KV HBM was possible at 2.1GHz. The analysis is validated by simulation for 2.1GHz LNA with ON chip Spiral inductors and ESD protection in Cadence Design Suite using IBM CSOI7TF 180nm RF Technology. The Overvoltage protection issues in Mixed Voltage IOs is addressed next. When two digital logic devices having different power supply levels are coupled an interface is generally required to prevent damage to transistors in the device having the lower power supply level. In the mixed-voltage I/O buffers, that interface t

Abstract

Robot localization is the problem of estimating a robot’s coordinates relative to an external reference frame, say a map. Given a map of its environment, the robot has to localize itself relative to this map by consulting its sensor data. The task of the robot is to find out where it is, through sensing and motion. Localization has proved to be the most critical aspect of mobile robot navigation. There are different forms of the localization problem: Position tracking, which seeks to compensate small dead reckoning errors (odometry errors) under the assumption that the initial robot pose is known, and global self- localization, which addresses the problem of localization with no apriori information i.e. a robot is not told its initial pose but instead has to determine it from scratch. In environments which possess relatively few features that enable a robot to unambiguously deter- mine its location, global localization algorithms can result in multiple hypotheses about the location of a robot. This is inevitable as the local environment seen by a robot repeats at several parts of the map. Such a situation occurs commonly in indoor navigation such as in corridors and office rooms. For effec- tive localization the robot has to be actively guided to those locations where there is a maximum chance of eliminating most of the ambiguous states. This task is often referred to as ‘active localization’. Active localization is based on the decision of ‘where to move’ and ‘where to look’ so as to best localize the robot. These actions mainly comprise of moving towards unique features in the map (e.g., distinct obsta- cles in a static environment), such that the sensor readings become unique and hence the robot localizes. This thesis presents a learning framework for implementing active localization. In a given map, the framework should be able to determine what actions to execute in a state, such that localization is ac- celerated. Since global localization poses the problem of perceptual aliasing, a belief state instead of a state is used to describe a robot’s position in the map. A belief-state MDP (Markov Decision Process) is used to capture the mapping between the robot’s belief state(s) and action(s) which will maximize the expected discounted rewards. The novelty of this work lies in the fact that it proposes a hierarchical belief state framework for the MDP to perform the active localization task. Markov localization is used typically to estimate a belief-state from a given set of robot’s sensor readings. This belief-state space being extensive is represented by a yet another hierarchical belief-state space, which makes the MDP more tractable. Hence the framework enables to perform localization with accuracy of fine grid granu- larity ( as in Markov localization) without having to handle large state space MDP(s). A mathematical formulation in detail for the framework has been provided. The feasibility of this framework is shown by testing it for different aspects of active localization task . It is used to perform multi-robot active localization, where the framework not only learns to seek out unique features in the map, but also learns to detect other robots, thereby facilitating localization even in the absence of map features. It is observed that multi-robot localization algorithm proves to be an intuitive, semi-distributed algorithm which is able to attain an autonomy with respect to the number of robots. That is to say, unlike a centralized algorithm that requires re-calculation from the scratch with the change in the number of agents involved, this algorithm can be re-used without re-training for a range of number of robots. The analysis of localization performance is done by comparing it with a learning framework which uses random walk. The framework is also used to perform transfer of the learned localization policy from a robot with higher sensor capabilities to the one with the lower sensor capabilities, thereby enabling the possibil- ity of accelerating localization in a setup with heterogenous robots. The change in the sensor range is considered as the change in observation function in MDP context and a transfer policy is designed to transfer the knowledge between the two belief states MDP(s). The localization performance with transfer is analyzed by comparing with performance , when learning is done from the scratch without any transfer. It is observed that transfer leads to a better learning rate as the performance reaches an optimum in lower number of training iterations. One of the observation also shows the effect of map type on transfer and suggests the fact that the advantages of transfer is more pre-dominant in the case of sparse maps than the dense maps. The contribution here lies in attempting transfer across belief states MDP(s), since transfer has been limited to just MDP(s) till now. Also the transfer aspect is