This paper addresses the problem of noisy Generalized Binary Search (GBS). GBS is a well-known greedy algorithm for determining a binary-valued hypothesis through a sequence of strategically selected queries. At each step, a query is selected that most evenly splits the hypotheses under consideration into two disjoint subsets, a natural generalization of the idea underlying classic binary search. GBS is used in many applications, including fault testing, machine diagnostics, disease diagnosis, job scheduling, image processing, computer vision, and active learning. In most of these cases, the responses to queries can be noisy. Past work has provided a partial characterization of GBS, but existing noise-tolerant versions of GBS are suboptimal in terms of sample complexity. This paper presents the first optimal algorithm for noisy GBS and demonstrates its application to learning multidimensional threshold functions.

Soft goals extend the classical model of planning with a simple model of preferences. The best plans are then not the ones with least cost but the ones with maximum utility, where the utility of a plan is the sum of the utilities of the soft goals achieved minus the plan cost. Finding plans with high utility appears to involve two linked problems: choosing a subset of soft goals to achieve and finding a low-cost plan to achieve them. New search algorithms and heuristics have been developed for planning with soft goals, and a new track has been introduced in the International Planning Competition (IPC) to test their performance. In this note, we show however that these extensions are not needed: soft goals do not increase the expressive power of the basic model of planning with action costs, as they can easily be compiled away. We apply this compilation to the problems of the net-benefit track of the most recent IPC, and show that optimal and satisficing cost-based planners do better on the compiled problems than optimal and satisficing net-benefit planners on the original problems with explicit soft goals. Furthermore, we show that penalties, or negative preferences expressing conditions to avoid, can also be compiled away using a similar idea.

There has been a lot of recent work on Bayesian methods for reinforcement learning exhibiting near-optimal online performance. The main obstacle facing such methods is that in most problems of interest, the optimal solution involves planning in an infinitely large tree. However, it is possible to obtain stochastic lower and upper bounds on the value of each tree node. This enables us to use stochastic branch and bound algorithms to search the tree efficiently. This paper proposes two such algorithms and examines their complexity in this setting.

Real-time heuristic search algorithms are suitable for situated agents that need to make their decisions in constant time. Since the original work by Korf nearly two decades ago, numerous extensions have been suggested. One of the most intriguing extensions is the idea of backtracking wherein the agent decides to return to a previously visited state as opposed to moving forward greedily. This idea has been empirically shown to have a significant impact on various performance measures. The studies have been carried out in particular empirical testbeds with specific real-time search algorithms that use backtracking. Consequently, the extent to which the trends observed are characteristic of backtracking in general is unclear. In this paper, we present the first entirely theoretical study of backtracking in real-time heuristic search. In particular, we present upper bounds on the solution cost exponential and linear in a parameter regulating the amount of backtracking. The results hold for a wide class of real-time heuristic search algorithms that includes many existing algorithms as a small subclass.

We present an approach to identifying concurrent patterns of behavior in in-sample temporal factor training data that precede Events of Interest (EoIs). We also present how to use discovered patterns to forecast EoIs in out-of-sample test data. The forecasting methodology is based on matching entities' observed behaviors to patterns discovered in retrospective data. This pattern concept is a generalization of previous pattern definitions. The new pattern concept, based around patterns observed in trends of factor data is based on a finite-state model where observed, sustained trends in a factor map to pattern states. Discovered patterns can be used as a diagnostic tool to better understand the dynamic conditions leading up to specific Event of Interest occurrences and hint at underlying causal structures leading to onsets and terminations of socio-political violence. We present a computationally efficient data-mining method to discover trend patterns. We give an example of using our pattern forecasting methodology to correctly forecast the advent and cessation of ethnic-religious violence in nation states with a low false-alarm rate.

We present the first class of mathematically rigorous, general, fully self-referential, self-improving, optimally efficient problem solvers. Inspired by Kurt Goedel's celebrated self-referential formulas (1931), such a problem solver rewrites any part of its own code as soon as it has found a proof that the rewrite is useful, where the problem-dependent utility function and the hardware and the entire initial code are described by axioms encoded in an initial proof searcher which is also part of the initial code. The searcher systematically and efficiently tests computable proof techniques (programs whose outputs are proofs) until it finds a provably useful, computable self-rewrite. We show that such a self-rewrite is globally optimal - no local maxima! - since the code first had to prove that it is not useful to continue the proof search for alternative self-rewrites. Unlike previous non-self-referential methods based on hardwired proof searchers, ours not only boasts an optimal order of complexity but can optimally reduce any slowdowns hidden by the O()-notation, provided the utility of such speed-ups is provable at all.

Did natural consciousness and intelligent systems arise out of a path that was co-evolutionary to evolution? Can we explain human self-consciousness as having risen out of such an evolutionary path? If so how could it have been? In this first part of a two-part paper (titled IXI), we take a learning system perspective to the problem of consciousness and intelligent systems, an approach that may look unseasonable in this age of fMRI's and high tech neuroscience. We posit conscious intelligent systems in natural environments and wonder how natural factors influence their design paths. Such a perspective allows us to explain seamlessly a variety of natural factors, factors ranging from the rise and presence of the human mind, man's sense of I, his self-consciousness and his looping thought processes to factors like reproduction, incubation, extinction, sleep, the richness of natural behavior, etc. It even allows us to speculate on a possible human evolution scenario and other natural phenomena.

INRA Toulouse, France Abstract Weighted Max-SA T is the optimization version of SA T and many important problems can be naturally encoded as such. Solving weighted Max-SA T is an important problem from both a theoretical and a practical point of view. In recent ye ars, there has been considerable interest in finding efficient solving techniques. Most of thi s work focus on the computation of good quality lower bounds to be used within a branch and bou nd DPLL-like algorithm. Most often, these lower bounds are described in a procedural way. Because of that, it is difficult to realize the logic that is behind. In this paper we introduce an original framework for Max-SA T that stresses the parallelism with classical SA T. Then, we extend the two basic SA T s olving techniques: search and inference. We show that many algorithmic tricks used in state-of-the-art Max-SA T solvers are easily expressable in logic terms with our framework in a unified manner. Besides, we introduce an original search algorithm that per forms a restricted amount of weighted resolution at each visited node. We empirically compare our algorithm w ith a variety of solving alternatives on several benchmarks. Our experiments, which constitute to the best of our knowledge the most comprehensive Max-sat eva luation ever reported, show that our algorithm is generally orders of magnitude faster t han any competitor. Preprint submitted to Elsevier Science 11 September 2018 1 Introduction Weighted Max-SA T is the optimization version of the SA T prob lem and many important problems can be naturally expressed as such. In recent years, there has been a considerable effort in finding efficient exact algorithms. A common drawback of all these alg orithms is that albeit the close relationship between SA T and Max-SA T, they cannot be easily described with logic terminology. For instance, the contributions of [11,12,13,14] are good quality lower bounds to be incorporated into a depth-first branch and bound procedure. These lower bounds are mostly defined in a procedural way and it is very difficult to see the logic that is behind the execution of the procedure. This is in contrast with SA T algorithms where the solving process can b e easily decomposed into atomic logical steps. In this paper we introduce an original framework for (weight ed) Max-SA T in which the notions of upper and lower bound are incorporated into the problem definition. Under this framework classical SA T is just a particular case of Max-SA T, and the main SA T solving techniques can be naturally extended. In pa rticular, we extend the basic simplification rules (for example, idempotency, absorption, unit clause reduction, etc) and introduce a new one, hardening, that does not make sense in the SA T context.

In this paper we propose a Multi-Objective Ant Colony Optimization (MOACO) algorithm called CHAC, which has been designed to solve the problem of finding the path on a map (corresponding to a simulated battlefield) that minimizes resources while maximizing safety. CHAC has been tested with two different state transition rules: an aggregative function that combines the heuristic and pheromone information of both objectives and a second one that is based on the dominance concept of multiobjective optimization problems. These rules have been evaluated in several different situations (maps with different degree of difficulty), and we have found that they yield better results than a greedy algorithm (taken as baseline) in addition to a military behaviour that is also better in the tactical sense. The aggregative function, in general, yields better results than the one based on dominance.

An approach to the classification problem of machine learning, based on building local classification rules, is developed. The local rules are considered as projections of the global classification rules to the event we want to classify. A massive global optimization algorithm is used for optimization of quality criterion. The algorithm, which has polynomial complexity in typical case, is used to find all high--quality local rules. The other distinctive feature of the algorithm is the integration of attributes levels selection (for ordered attributes) with rules searching and original conflicting rules resolution strategy. The algorithm is practical; it was tested on a number of data sets from UCI repository, and a comparison with the other predicting techniques is presented.