Other Papers
2023
Segev Wasserkrug, Takayuki Osogami, Who Benefits from a Multi-Cloud Market? A Trading Networks Based Analysis, arXiv:2310.12666, 2023.
2017
Takayuki Osogami, Boltzmann machines for time-series, IBM Research Report, RT0980, 2017. PDF
Abstract We review Boltzmann machines extended for time-series. These models often have recurrent structure, and back propagration through time (BPTT) is used to learn their parameters. The per-step computational complexity of BPTT in online learning, however, grows linearly with respect to the length of preceding time-series (i.e., learning rule is not local in time), which limits the applicability of BPTT in online learning. We then review dynamic Boltzmann machines (DyBMs), whose learning rule is local in time. DyBM's learning rule relates to spike-timing dependent plasticity (STDP), which has been postulated and experimentally confirmed for biological neural networks.
Takayuki Osogami, Boltzmann machines and energy-based models, IBM Research Report, RT0979, 2017. PDF
Abstract We review Boltzmann machines and energy-based models. A Boltzmann machine defines a probability distribution over binary-valued patterns. One can learn parameters of a Boltzmann machine via gradient based approaches in a way that log likelihood of data is increased. The gradient and Laplacian of a Boltzmann machine admit beautiful mathematical representations, although computing them is in general intractable. This intractability motivates approximate methods, including Gibbs sampler and contrastive divergence, and tractable alternatives, namely energy-based models.
2015
Takayuki Osogami and MakotoOtsuka, Learning dynamic Boltzmann machines with spike-timing dependent plasticity, Technical Report RT0967, IBM Research, 2015. PDF
Abstract We propose a particularly structured Boltzmann machine, which we refer to as a dynamic Boltzmann machine (DyBM), as a stochastic model of a multidimensional time-series. The DyBM can have infinitely many layers of units but allows exact and efficient inference and learning when its parameters have a proposed structure. This proposed structure is motivated by postulates and observations, from biological neural networks, that the synaptic weight is strengthened or weakened, depending on the timing of spikes (i.e., spike-timing dependent plasticity or STDP). We show that the learning rule of updating the parameters of the DyBM in the direction of maximizing the likelihood of given time-series can be interpreted as STDP with long term potentiation and long term depression. The learning rule has a guarantee of convergence and can be performed in a distributed matter (i.e., local in space) with limited memory (i.e., local in time).
Takayuki Osogami and Tetsuro Morimura, When the optimal policy is independent of the initial state, IBM Research Report, RT0966, 2015. PDF
Abstract A Markov decision process (MDP) is a popular model of sequential decision making, but its standard objective of minimizing cumulative cost is often inadequate, for example, to avoid the possibility of large loss. Risk-sensitive objective functions and constraints have thus been proposed for MDPs. Unlike the standard MDP, however, the optimal policy for some of these MDPs can depend on the initial states, so that the optimal policy can change over time. We show that an agent can surely incur larger cumulative cost by following the latest optimal policy at every state than by following other policies. We then establish sufficient conditions on the objective function and on the constraints for the optimal policies to be consistent between the initial states. We also show when the sufficient conditions are necessary. We discuss implications of our results to the MDPs that have been studied in the literature, stating whether their optimal policies depend on the initial states.
2012
Takayuki Osogami, Takashi Imamichi, Hideyuki Mizuta, Tetsuro Morimura, Rudy Raymond, Toyotaro Suzumura, Rikiya Takahashi, and Tsuyoshi Ide, IBM Mega Traffic Simulator, Technical Report, IBM Research, RT0896, December 2012. PDF
Abstract We propose a model of design processes, which we refer to as a rework structure matrix (RSM). An RSM is superior to the design structure matrix (DSM), which has been a popular model for analyzing design processes, in that an RSM allows one to quickly optimize the design processes so that rework is minimized. We show that the design process can be optimized under an RSM by solving a minimum feedback arc set problem, which has been studied extensively in the literature and for which various algorithms have been developed.
Takayuki Osogami, "Robustness of time-consistent Markov decision processes", The 15th Information-based Induction Science Workshop (IBIS 2012)
2010
Takayuki Osogami, Overcoming limitations of expected utility with iterated risk measures, Technical Report, IBM Research, RT0921, November 2010. PDF
2007
Takayuki Osogami et al., "Optimizing Design Processes via Rework Structure Matrices," in Proceedings of Scheduling Symposium 2007, Kyoto, Japan, September 2007.
Abstract We propose a model of design processes, which we refer to as a rework structure matrix (RSM). An RSM is superior to the design structure matrix (DSM), which has been a popular model for analyzing design processes, in that an RSM allows one to quickly optimize the design processes so that rework is minimized. We show that the design process can be optimized under an RSM by solving a minimum feedback arc set problem, which has been studied extensively in the literature and for which various algorithms have been developed.
2006
[NDRM06] Rikiya Takahashi and Takayuki Osogami, "Action-sensitive hidden Markov model of customer behavior," in Proceedings of Tsukuba-Tohoku Joint Workshop on New Directions of Research in Marketing, Tsukuba, Japan, December 2006.
Abstract To maximize a firm's long-term cumulative profits from direct marketing, decisions need to be made based on models of customer behavior which incorporate both the short-term and long-term effects of marketing actions on customer behavior. We propose a hidden Markov model (HMM) of customer behavior, which we refer to as the action-sensitive HMM (AHMM). In AHMM, it is assumed that marketing actions not only can affect immediate purchases of a customer but also can affect how a customer transitions between states, which reflects the long-term effects of the marketing actions. AHMM can easily be extended to a Markov decision process (MDP) for optimizing marketing actions so that the long-term cumulative profits are maximized. Another advantage of AHMM over existing HMMs of customer behavior is that AHMM can capture more detailed properties of the inter-purchase times of a customer without increasing the number of customer states. We also propose an algorithm for estimating the parameters of the AHMM, which cannot be estimated by standard approaches. The key idea in our algorithm is that it first estimates the transition probabilities between pairs of customer states and marketing actions. The transition probabilities of the state-action pairs are then used to estimate transition probabilities between customer states conditioned on marketing actions.
2004
Takayuki Osogami, Analysis of a QBD Process that Depends on Background QBD Processes, Technical Report CMU-CS-04-163, 2004. PDF
Abstract We defi.ne a class of Markov chains that are called recursive foreground-background quasi-birth-and-death (RFBQBD) processes, and describe approximate analyses of the RFBQBD process. We show that the RFBQBD process can model many interesting multiserver systems with multiple classes of jobs, where the behavior of certain classes of jobs has inherent dependencies on the behavior of other classes of jobs. Many of these multiserver systems cannot be analyzed via traditional approaches. We also evaluate the running time and accuracy of our analyses numerically by applying them to the performance analysis of a particular task assignment policy in a multiserver system. Our numerical evaluation shows that the error in our approximations is within 5% in the mean queue length and within 10% in the second moment of the queue length distribution for a range of loads and job size distributions.
1998
Takayuki Osogami, Approaches to 3D Free Form Cutting and Packing Problems and Their Applications: A Survey, IBM TRL Research Report, RT0287, 1998. PDF
Abstract This survey paper describes approaches to and applications of 3D free-form cutting and packing problems. Some initial work on these problems has been done in the last few years. The approaches used, which include simulated annealing, the genetic algorithm, and heuristics with simulated annealing, are described, and applications of the problems are introduced. They include 3D product layout and rapid prototyping, which is a new technology for quickly creating 3D products with computer-aided design systems. Finally, practical aspects of the problems involved in rapid prototyping
are discussed.