arXiv subjects

Huizhen Yu

Publications and source records attributed to Huizhen Yu.

24 records · Page 2Linked to original sources

arXiv2015

Emphatic Temporal-Difference Learning

Emphatic algorithms are temporal-difference learning algorithms that change their effective state distribution by selectively emphasizing and de-emphasizing their updates on different time steps. Recent works by Sutton, Mahmood and White (2015), and Yu (2015) show that by varying the emphasis in a particular way, these algorithms become stable and convergent under off-policy training with linear function approximation. This paper serves as a unified summary of the available results from both works. In addition, we demonstrate the empirical benefits from the flexibility of emphatic algorithms, including state-dependent discounting, state-dependent bootstrapping, and the user-specified allocation of function approximation resources.

cs.LG↗

arXiv2015

On Convergence of Value Iteration for a Class of Total Cost Markov Decision Processes

We consider a general class of total cost Markov decision processes (MDP) in which the one-stage costs can have arbitrary signs, but the sum of the negative parts of the one-stage costs is finite for all policies and all initial states. We refer to this class as the General Convergence (GC for short) total cost model, and we study the convergence of value iteration for the GC model, in the Borel MDP framework with universally measurable policies. Our main results include: (i) convergence of value iteration when starting from certain functions above the optimal cost function; (ii) convergence of transfinite value iteration starting from zero, in the special case where the optimal cost function is nonnegative; and (iii) partial convergence of value iteration starting from zero, for a subset of initial states. These results extend several previously known results about the convergence of value iteration for either positive costs problems or GC total cost problems. In particular, the first result on convergence of value iteration from above extends a theorem of van der Wal for the GC model. The second result relates to Maitra and Sudderth's analysis of transfinite value iteration for the positive costs model. It suggests connections between the two total cost models when the optimal cost function is nonnegative, and it leads to additional results on the convergence of ordinary non-transfinite value iteration, with a suitably defined dynamic programming operator, for finite state or finite control GC problems. The third result on partial convergence of value iteration is motivated by Whittle's bridging condition for the positive costs model, and provides a novel extension of the bridging condition to the GC model, where there are no sign constraints on the costs.

math.OC↗

arXiv2014

Stochastic Shortest Path Games and Q-Learning

We consider a class of two-player zero-sum stochastic games with finite state and compact control spaces, which we call stochastic shortest path (SSP) games. They are undiscounted total cost stochastic dynamic games that have a cost-free termination state. Exploiting the close connection of these games to single-player SSP problems, we introduce novel model conditions under which we show that the SSP games have strong optimality properties, including the existence of a unique solution to the dynamic programming equation, the existence of optimal stationary policies, and the convergence of value and policy iteration. We then focus on finite state and control SSP games and the classical Q-learning algorithm for computing the value function. Q-learning is a model-free, asynchronous stochastic iterative algorithm. By the theory of stochastic approximation involving monotone nonexpansive mappings, it is known to converge when its associated dynamic programming equation has a unique solution and its iterates are bounded with probability one. For the SSP case, as the main result of this paper, we prove the boundedness of the Q-learning iterates under our proposed model conditions, thereby establishing completely the convergence of Q-learning for a broad class of total cost finite-space stochastic games.

math.OC↗

arXiv2014

A Mixed Value and Policy Iteration Method for Stochastic Control with Universally Measurable Policies

We consider stochastic control models with Borel spaces and universally measurable policies. For such models the standard policy iteration is known to have difficult measurability issues and cannot be carried out in general. We present a mixed value and policy iteration method that circumvents this difficulty. The method allows the use of stationary policies in computing the optimal cost function, in a manner that resembles policy iteration. It can also be used to address similar difficulties of policy iteration in the context of upper and lower semicontinuous models. We analyze the convergence of the method in infinite horizon total cost problems, for the discounted case where the one-stage costs are bounded, and for the undiscounted case where the one-stage costs are nonpositive or nonnegative. For undiscounted total cost problems with nonnegative one-stage costs, we also give a new convergence theorem for value iteration, which shows that value iteration converges whenever it is initialized with a function that is above the optimal cost function and yet bounded by a multiple of the optimal cost function. This condition resembles Whittle's bridging condition and is partly motivated by it. The theorem is also partly motivated by a result of Maitra and Sudderth, which showed that value iteration, when initialized with the constant function zero, could require a transfinite number of iterations to converge. We use the new convergence theorem for value iteration to establish the convergence of our mixed value and policy iteration method for the nonnegative cost case.

math.OC↗

arXiv2012

Discretized Approximations for POMDP with Average Cost

In this paper, we propose a new lower approximation scheme for POMDP with discounted and average cost criterion. The approximating functions are determined by their values at a finite number of belief points, and can be computed efficiently using value iteration algorithms for finite-state MDP. While for discounted problems several lower approximation schemes have been proposed earlier, ours seems the first of its kind for average cost problems. We focus primarily on the average cost case, and we show that the corresponding approximation can be computed efficiently using multi-chain algorithms for finite-state MDP. We give a preliminary analysis showing that regardless of the existence of the optimal average cost J in the POMDP, the approximation obtained is a lower bound of the liminf optimal average cost function, and can also be used to calculate an upper bound on the limsup optimal average cost function, as well as bounds on the cost of executing the stationary policy associated with the approximation. Weshow the convergence of the cost approximation, when the optimal average cost is constant and the optimal differential cost is continuous.

cs.AI↗

arXiv2012

A Function Approximation Approach to Estimation of Policy Gradient for POMDP with Structured Policies

We consider the estimation of the policy gradient in partially observable Markov decision processes (POMDP) with a special class of structured policies that are finite-state controllers. We show that the gradient estimation can be done in the Actor-Critic framework, by making the critic compute a "value" function that does not depend on the states of POMDP. This function is the conditional mean of the true value function that depends on the states. We show that the critic can be implemented using temporal difference (TD) methods with linear function approximations, and the analytical results on TD and Actor-Critic can be transfered to this case. Although Actor-Critic algorithms have been used extensively in Markov decision processes (MDP), up to now they have not been proposed for POMDP as an alternative to the earlier proposal GPOMDP algorithm, an actor-only method. Furthermore, we show that the same idea applies to semi-Markov problems with a subset of finite-state controllers.

cs.LG↗