etd@IISc Collection:
http://hdl.handle.net/2005/18
Tue, 21 Nov 2017 06:33:29 GMT2017-11-21T06:33:29ZFeature Adaptation Algorithms for Reinforcement Learning with Applications to Wireless Sensor Networks And Road Traffic Control
http://hdl.handle.net/2005/2664
Title: Feature Adaptation Algorithms for Reinforcement Learning with Applications to Wireless Sensor Networks And Road Traffic Control
Authors: Prabuchandran, K J
Abstract: Many sequential decision making problems under uncertainty arising in engineering, science and economics are often modelled as Markov Decision Processes (MDPs). In the setting of MDPs, the goal is to and a state dependent optimal sequence of actions that minimizes a certain long-term performance criterion. The standard dynamic programming approach to solve an MDP for the optimal decisions requires a complete model of the MDP and is computationally feasible only for small state-action MDPs. Reinforcement learning (RL) methods, on the other hand, are model-free simulation based approaches for solving MDPs. In many real world applications, one is often faced with MDPs that have large state-action spaces whose model is unknown, however, whose outcomes can be simulated. In order to solve such (large) MDPs, one either resorts to the technique of function approximation in conjunction with RL methods or develops application specific RL methods. A solution based on RL methods with function approximation comes with the associated problem of choosing the right features for approximation and a solution based on application specific RL methods primarily relies on utilizing the problem structure. In this thesis, we investigate the problem of choosing the right features for RL methods based on function approximation as well as develop novel RL algorithms that adaptively obtain best features for approximation. Subsequently, we also develop problem specie RL methods for applications arising in the areas of wireless sensor networks and road traffic control.
In the first part of the thesis, we consider the problem of finding the best features for value function approximation in reinforcement learning for the long-run discounted cost objective. We quantify the error in the approximation for any given feature and the approximation parameter by the mean square Bellman error (MSBE) objective and develop an online algorithm to optimize MSBE.
Subsequently, we propose the first online actor-critic scheme with adaptive bases to find a locally optimal (control) policy for an MDP under the weighted discounted cost objective. The actor performs gradient search in the space of policy parameters using simultaneous perturbation stochastic approximation (SPSA) gradient estimates. This gradient computation however requires estimates of the value function of the policy. The value function is approximated using a linear architecture and its estimate is obtained from the critic. The error in approximation of the value function, however, results in sub-optimal policies. Thus, we obtain the best features by performing a gradient descent on the Grassmannian of features to minimize a MSBE objective. We provide a proof of convergence of our control algorithm to a locally optimal policy and show numerical results illustrating the performance of our algorithm.
In our next work, we develop an online actor-critic control algorithm with adaptive feature tuning for MDPs under the long-run average cost objective. In this setting, a gradient search in the policy parameters is performed using policy gradient estimates to improve the performance of the actor. The computation of the aforementioned gradient however requires estimates of the differential value function of the policy. In order to obtain good estimates of the differential value function, the critic adaptively tunes the features to obtain the best representation of the value function using gradient search in the Grassmannian of features. We prove that our actor-critic algorithm converges to a locally optimal policy. Experiments on two different MDP settings show performance improvements resulting from our feature adaptation scheme.
In the second part of the thesis, we develop problem specific RL solution methods for the two aforementioned applications. In both the applications, the size of the state-action space in the formulated MDPs is large. However, by utilizing the problem structure we develop scalable RL algorithms.
In the wireless sensor networks application, we develop RL algorithms to find optimal energy management policies (EMPs) for energy harvesting (EH) sensor nodes. First, we consider the case of a single EH sensor node and formulate the problem of finding an optimal EMP in the discounted cost MDP setting. We then propose two RL algorithms to maximize network performance. Through simulations, our algorithms are seen to outperform the algorithms in the literature. Our RL algorithms for the single EH sensor node do not scale when there are multiple sensor nodes. In our second work, we consider the problem of finding optimal energy sharing policies that maximize the network performance of a system comprising of multiple sensor nodes and a single energy harvesting (EH) source. We develop efficient energy sharing algorithms, namely Q-learning algorithm with exploration mechanisms based on the -greedy method as well as upper confidence bound (UCB). We extend these algorithms by incorporating state and action space aggregation to tackle state-action space explosion in the MDP. We also develop a cross entropy based method that incorporates policy parameterization in order to find near optimal energy sharing policies. Through numerical experiments, we show that our algorithms yield energy sharing policies that outperform the heuristic greedy method.
In the context of road traffic control, optimal control of traffic lights at junctions or traffic signal control (TSC) is essential for reducing the average delay experienced by the road users. This problem is hard to solve when simultaneously considering all the junctions in the road network. So, we propose a decentralized multi-agent reinforcement learning (MARL) algorithm for solving this problem by considering each junction in the road network as a separate agent (controller) to obtain dynamic TSC policies. We propose two approaches to minimize the average delay. In the first approach, each agent decides the signal duration of its phases in a round-robin (RR) manner using the multi-agent Q-learning algorithm. We show through simulations over VISSIM (microscopic traffic simulator) that our round-robin MARL algorithms perform significantly better than both the standard fixed signal timing (FST) algorithm and the saturation balancing (SAT) algorithm over two real road networks. In the second approach, instead of optimizing green light duration, each agent optimizes the order of the phase sequence. We then employ our MARL algorithms by suitably changing the state-action space and cost structure of the MDP. We show through simulations over VISSIM that our non-round robin MARL algorithms perform significantly better than the FST, SAT and the round-robin MARL algorithms based on the first approach. However, on the other hand, our round-robin MARL algorithms are more practically viable as they conform with the psychology of road users.Tue, 19 Sep 2017 18:30:00 GMThttp://hdl.handle.net/2005/26642017-09-19T18:30:00ZRetweet Profiling - Study Dissemination of Twitter Messages
http://hdl.handle.net/2005/2668
Title: Retweet Profiling - Study Dissemination of Twitter Messages
Authors: Rangnani, Soniya
Abstract: Social media has become an important means of everyday communication. It is a mechanism for “sharing” and “resharing” of information. While social network platforms provide the means to users for resharing/reblogging (aka retweeting), it remains unclear what motivates users to share. Predicting the spread of content is quite important for several purposes such as viral marketing, popular news detection, personalized message recommendation and on-line advertisement. Social content systems store all the information produced in the interactions between users. However, to turn this data into information that allows us to extract patterns, it is important to consider the different phenomena involved in these interactions. In this work, two phenomena that influence the evolution of networks are studied for Twitter: diffusion of information and communication among users.
Previous studies have shown that history of interaction among users and properties of the message are good attributes to understand the retweet behavior of users. Factors like content of message and time are less investigated. We propose a prediction model for retweet actions of users. It formulates a function which ranks the users according to how receptive they are to a particular message. The function generates a confidence score for the edges joining the initiator of the message and the followers. Two different pieces of information propagate through different users in the network. We divide the task of calculating confidence score into two parts. The first part is independent of the test tweet. It models transmission rate of the tie between the initiator and the follower. We call this as ‘Pairwise Influence Estimation’. The second part incorporates the tweet properties and user activeness as per time in the ranking function. The proposed model exploits all the dimensions of information dif-fusion process-influence, content and temporal properties. We have captured local aspects of diffusion.
It has been observed that users do not read all the messages on their site. This results in shortcomings in the above models. Considering this, we first study the temporal behavior of users’ activities, which directly reflects their availability pertaining to the upcoming post. Also, as it is a continuous task of predicting retweet behavior, we design a user-centric, and temporally localized incremental classification model by considering the fact that users do not read all their tweets. We have tested the effectiveness of this model by using real data from Twitter. We demonstrate that the new proposed model is more accurate in describing the information propagation in microblog compared to the existing methods. Our model works well when we consider different classes of users depending on their activity patterns. In addition, we also investigate the parameters of the model for different classes of users. We report some interesting distinguishing patterns in retweeting behavior of users.Fri, 22 Sep 2017 18:30:00 GMThttp://hdl.handle.net/2005/26682017-09-22T18:30:00ZPlan Bouquets : An Exploratory Approach to Robust Query Processing
http://hdl.handle.net/2005/2686
Title: Plan Bouquets : An Exploratory Approach to Robust Query Processing
Authors: Dutt, Anshuman
Abstract: Over the last four decades, relational database systems, with their mathematical basis in first-order logic, have provided a congenial and efficient environment to handle enterprise data during its entire life cycle of generation, storage, maintenance and processing. An organic reason for their pervasive popularity is intrinsic support for declarative user queries, wherein the user only specifies the end objectives, and the system takes on the responsibility of identifying the most efficient means, called “plans”, to achieve these objectives. A crucial input to generating efficient query execution plans are the compile-time estimates of the data volumes that are output by the operators implementing the algebraic predicates present in the query. These volume estimates are typically computed using the “selectivities” of the predicates. Unfortunately, a pervasive problem encountered in practice is that these selectivities often differ significantly from the values actually encountered during query execution, leading to poor plan choices and grossly inflated response times. While the database research community has spent considerable efforts to address the above challenge, the prior techniques all suffer from a systemic limitation - the inability to provide any guarantees on the execution performance.
In this thesis, we materially address this long-standing open problem by developing a radically different query processing strategy that lends itself to attractive guarantees on run-time performance. Specifically, in our approach, the compile-time estimation process is completely eschewed for error-prone selectivities. Instead, from the set of optimal plans in the query’s selectivity error space, a limited subset called the “plan bouquet”, is selected such that at least one of the bouquet plans is 2-optimal at each location in the space. Then, at run time, an exploratory sequence of cost-budgeted executions from the plan bouquet is carried out, eventually finding a plan that executes to completion within its assigned budget. The duration and switching of these executions is controlled by a graded progression of isosurfaces projected onto the optimal performance profile. We prove that this construction provides viable guarantees on the worst-case performance relative to an oracular system that magically possesses accurate apriori knowledge of all selectivities. Moreover, it ensures repeatable execution strategies across different invocations of a query, an extremely desirable feature in industrial settings.
Our second contribution is a suite of techniques that substantively improve on the performance guarantees offered by the basic bouquet algorithm. First, we present an algorithm that skips carefully chosen executions from the basic plan bouquet sequence, leveraging the observation that an expensive execution may provide better coverage as compared to a series of cheaper siblings, thereby reducing the aggregate exploratory overheads. Next, we explore randomized variants with regard to both the sequence of plan executions and the constitution of the plan bouquet, and show that the resulting guarantees are markedly superior, in expectation, to the corresponding worst case values.
From a deployment perspective, the above techniques are appealing since they are completely “black-box”, that is, non-invasive with regard to the database engine, implementable using only API features that are commonly available in modern systems. As a proof of concept, the bouquet approach has been fully prototyped in QUEST, a Java-based tool that provides a visual and interactive demonstration of the bouquet identification and execution phases. In similar spirit, we propose an efficient isosurface identification algorithm that avoids exploration of large portions of the error space and drastically reduces the effort involved in bouquet construction.
The plan bouquet approach is ideally suited for “canned” query environments, where the computational investment in bouquet identification is amortized over multiple query invocations. The final contribution of this thesis is extending the advantage of compile-time sub-optimality guarantees to ad hoc query environments where the overheads of the off-line bouquet identification may turn out to be impractical. Specifically, we propose a completely revamped bouquet algorithm that constructs the cost-budgeted execution sequence in an “on-the-fly” manner. This is achieved through a “white-box” interaction style with the engine, whereby the plan output cardinalities exposed by the engine are used to compute lower bounds on the error-prone selectivities during plan executions. For this algorithm, the sub-optimality guarantees are in the form of a low order polynomial of the number of error-prone selectivities in the query.
The plan bouquet approach has been empirically evaluated on both PostgreSQL and a commercial engine ComOpt, over the TPC-H and TPC-DS benchmark environments. Our experimental results indicate that it delivers orders of magnitude improvements in the worst-case behavior, without impairing the average-case performance, as compared to the native optimizers of these systems. In absolute terms, the worst case sub-optimality is upper bounded by 20 across the suite of queries, and the average performance is empirically found to be within a factor of 4 wrt the optimal. Even with the on-the-fly bouquet algorithm, the guarantees are found to be within a factor of 3 as compared to those achievable in the corresponding canned query environment.
Overall, the plan bouquet approach provides novel performance guarantees that open up exciting possibilities for robust query processing.Mon, 25 Sep 2017 18:30:00 GMThttp://hdl.handle.net/2005/26862017-09-25T18:30:00ZProvable Methods for Non-negative Matrix Factorization
http://hdl.handle.net/2005/2739
Title: Provable Methods for Non-negative Matrix Factorization
Authors: Pani, Jagdeep
Abstract: Nonnegative matrix factorization (NMF) is an important data-analysis problem which concerns factoring a given d n matrix A with nonnegative entries into matrices B and C where B and C are d k and k n with nonnegative entries. It has numerous applications including Object recognition, Topic Modelling, Hyper-spectral imaging, Music transcription etc. In general, NMF is intractable and several heuristics exists to solve the problem of NMF. Recently there has been interest in investigating conditions under which NMF can be tractably recovered. We note that existing attempts make unrealistic assumptions and often the associated algorithms tend to be not scalable.
In this thesis, we make three major contributions: First, we formulate a model of NMF with assumptions which are natural and is a substantial weakening of separability. Unlike requiring a bound on the error in each column of (A BC) as was done in much of previous work, our assumptions are about aggregate errors, namely spectral norm of (A BC) i.e. jjA BCjj2 should be low. This is a much weaker error assumption and the associated B; C would be much more resilient than existing models. Second, we describe a robust polynomial time SVD-based algorithm, UTSVD, with realistic provable error guarantees and can handle higher levels of noise than previous algorithms. Indeed, experimentally we show that existing NMF models, which are based on separability assumptions, degrade much faster than UTSVD, in the presence of noise. Furthermore, when the data has dominant features, UTSVD significantly outperforms existing models. On real life datasets we again see a similar outperformance of UTSVD on clustering tasks. Finally, under a weaker model, we prove a robust version of uniqueness of NMF, where again, the word \robust" refers to realistic error bounds.Mon, 30 Oct 2017 18:30:00 GMThttp://hdl.handle.net/2005/27392017-10-30T18:30:00Z