An end-to-end system was created at Genworth Financial to automate the underwriting of long-term care (LTC) and life insurance applications. Relying heavily on artificial intelligence techniques, the system has been in production since December 2002 and in 2004 completely automates the underwriting of 19 percent of the LTC applications. A fuzzy logic rules engine encodes the underwriter guidelines and an evolutionary algorithm optimizes the engine's performance. Finally, a natural language parser is used to improve the coverage of the underwriting system.

This article is a sequel to an article titled "A New Direction in AI -- Toward a Computational Theory of Perceptions," which appeared in the Spring 2001 issue of AI Magazine (volume 22, No. 1, 73-84). The concept of precisiated natural language (PNL) was briefly introduced in that article, and PNL was employed as a basis for computation with perceptions. In what follows, the conceptual structure of PNL is described in greater detail, and PNL's role in knowledge representation, deduction, and concept definition is outlined and illustrated by examples. What should be understood is that PNL is in its initial stages of development and that the exposition that follows is an outline of the basic ideas that underlie PNL rather than a definitive theory. A natural language is basically a system for describing perceptions. Perceptions, such as perceptions of distance, height, weight, color, temperature, similarity, likelihood, relevance, and most other attributes of physical and mental objects are intrinsically imprecise, reflecting the bounded ability of sensory organs, and ultimately the brain, to resolve detail and store information. In this perspective, the imprecision of natural languages is a direct consequence of the imprecision of perceptions (Zadeh 1999, 2000). How can a natural language be precisiated -- precisiated in the sense of making it possible to treat propositions drawn from a natural language as objects of computation? This is what PNL attempts to do. In PNL, precisiation is accomplished through translation into what is termed a precisiation language. In the case of PNL, the precisiation language is the generalized-constraint language (GCL), a language whose elements are so-called generalized constraints and their combinations. What distinguishes GCL from languages such as Prolog, LISP, SQL, and, more generally, languages associated with various logical systems, for example, predicate logic, modal logic, and so on, is its much higher expressive power. The conceptual structure of PNL mirrors two fundamental facets of human cognition: (a) partiality and (b) granularity (Zadeh 1997). Partiality relates to the fact that most human concepts are not bivalent, that is, are a matter of degree. Thus, we have partial understanding, partial truth, partial possibility, partial certainty, partial similarity, and partial relevance, to cite a few examples. Similarly, granularity and granulation relate to clumping of values of attributes, forming granules with words as labels, for example, young, middle-aged, and old as labels of granules of age. Existing approaches to natural language processing are based on bivalent logic -- a logic in which shading of truth is not allowed. PNL abandons bivalence. By so doing, PNL frees itself from limitations imposed by bivalence and categoricity, and opens the door to new approaches for dealing with long-standing problems in AI and related fields (Novak 1991). At this juncture, PNL is in its initial stages of development. As it matures, PNL is likely to find a variety of applications, especially in the realms of world knowledge representation, concept definition, deduction, decision, search, and question answering.

Convergence for iterative reinforcement learning algorithms like TD(O) depends on the sampling strategy for the transitions. However, in practical applications it is convenient to take transition data from arbitrary sources without losing convergence. In this paper we investigate the problem of repeated synchronous updates based on a fixed set of transitions. Our main theorem yields sufficient conditions of convergence for combinations of reinforcement learning algorithms and linear function approximation. This allows to analyse if a certain reinforcement learning algorithm and a certain function approximator are compatible.

There are several reinforcement learning algorithms that yield approximate solutions for the problem of policy evaluation when the value function is represented with a linear function approximator. In this paper we show that each of the solutions is optimal with respect to a specific objective function.

We establish a new hardness result that shows that the difficulty of planning in factored Markov decision processes is representational rather than just computational. More precisely, we give a fixed family of factored MDPs with linear rewards whose optimal policies and value functions simply cannot be represented succinctly in any standard parametric form. Previous hardness results indicated that computing good policies from the MDP parameters was difficult, but left open the possibility of succinct function approximation for any fixed factored MDP. Our result applies even to policies which yield a polynomially poor approximation to the optimal value, and highlights interesting connections with the complexity class of Arthur-Merlin games.

Convergence for iterative reinforcement learning algorithms like TD(O) depends on the sampling strategy for the transitions. However, in practical applications it is convenient to take transition data from arbitrary sources without losing convergence. In this paper we investigate the problem of repeated synchronous updates based on a fixed set of transitions. Our main theorem yields sufficient conditions of convergence for combinations of reinforcement learning algorithms and linear function approximation. This allows to analyse if a certain reinforcement learning algorithm and a certain function approximator are compatible.

There are several reinforcement learning algorithms that yield approximate solutions for the problem of policy evaluation when the value function is represented with a linear function approximator. In this paper we show that each of the solutions is optimal with respect to a specific objective function.

We establish a new hardness result that shows that the difficulty of planning in factored Markov decision processes is representational rather than just computational. More precisely, we give a fixed family of factored MDPs with linear rewards whose optimal policies and value functions simply cannot be represented succinctly in any standard parametric form. Previous hardness results indicated that computing good policies from the MDP parameters was difficult, but left open the possibility of succinct function approximation for any fixed factored MDP. Our result applies even to policies which yield a polynomially poor approximation to the optimal value, and highlights interesting connections with the complexity class of Arthur-Merlin games.

Convergence for iterative reinforcement learning algorithms like TD(O) depends on the sampling strategy for the transitions. However, inpractical applications it is convenient to take transition data from arbitrary sources without losing convergence. In this paper we investigate the problem of repeated synchronous updates based on a fixed set of transitions. This allows to analyse if a certain reinforcement learning algorithm and a certain functionapproximator are compatible. For the combination of the residual gradient algorithm with grid-based linear interpolation we show that there exists a universal constant learning rate such that the iteration converges independently of the concrete transition data. 1 Introduction The strongest convergence guarantees for reinforcement learning (RL) algorithms are available for the tabular case, where temporal difference algorithms for both policy evaluation and the general control problem converge with probability one independently of the concrete sampling strategy as long as all states are sampled infinitely often and the learning rate is decreased appropriately [2].

There are several reinforcement learning algorithms that yield approximate solutionsfor the problem of policy evaluation when the value function is represented with a linear function approximator. In this paper we show that each of the solutions is optimal with respect to a specific objective function.