We examine techniques for combining generalized policies with search algorithms to exploit the strengths and overcome the weaknesses of each when solving probabilistic planning problems. The Action Schema Network (ASNet) is a recent contribution to planning that uses deep learning and neural networks to learn generalized policies for probabilistic planning problems. ASNets are well suited to problems where local knowledge of the environment can be exploited to improve performance, but may fail to generalize to problems they were not trained on. Monte-Carlo Tree Search (MCTS) is a forward-chaining state space search algorithm for optimal decision making which performs simulations to incrementally build a search tree and estimate the values of each state. Although MCTS can achieve state-of-the-art results when paired with domain-specific knowledge, without this knowledge, MCTS requires a large number of simulations in order to obtain reliable state-value estimates. By combining ASNets with MCTS, we are able to improve the capability of an ASNet to generalize beyond the distribution of problems it was trained on, as well as enhance the navigation of the search space by MCTS.

We present a novel search scheme for privacy-preserving multi-agent planning. Inspired by UCT search, the scheme is based on growing an asynchronous search tree by running repeated trials through the tree. We describe key differences to classical multi-agent forward search, discuss theoretical properties of the presented approach, and evaluate it based on benchmarks from the CoDMAP competition. As a secondary contribution, we describe a technique that extends the regular search approach by small explorative trials which are performed subsequent to each node expansion. We show that this technique significantly increases the number of problems solved for all algorithms considered, including MAFS.

My thesis is largely focused on the parallelisation of UCT (and other Best-First Search techniques) and the ramifications of doing so. I have identified issues with chunking in UCT, created by some forms of parallelisation, and developed a solution to this involving buffering of simulations that appear "out of order" and reevaluation of propagation data. I have developed a technique for scalable distribution of both tree data and computation across a large scale compute cluster. The context of most of my work is General Game Playing, but the techniques themselves are largely agnostic to domain.

When an optimal solution is not required, satisficing search methods such as greedy best-first search are often used to find solutions quickly. In work on satisficing search, there has been substantial attention devoted to how to solve problems associated with local minima or plateaus in the heuristic function. One technique that has been shown to be quite promising is using an alternative heuristic function that does not estimate cost-to-go, but rather estimates distance-to-go. There is currently little beyond intuition to explain its superiority. We begin by empirically showing that the success of the distance-to-go heuristic appears related to its having smaller local minima. We then discuss a reasonable theoretical model of heuristics and show that, under this model, the expected size of local minima is higher for a cost-to-go heuristic than a distance-to-go heuristic, offering a possible explanation as to why distance-to-go heuristics tend to outperform cost-to-go heuristics.

Most work in heuristic search considers problems where a low cost solution is preferred (MIN problems). In this paper, we investigate the complementary setting where a solution of high reward is preferred (MAX problems). Example MAX problems include finding a longest simple path in a graph, maximal coverage, and various constraint optimization problems. We examine several popular search algorithms for MIN problems and discover the curious ways in which they misbehave on MAX problems. We propose modifications that preserve the original intentions behind the algorithms but allow them to solve MAX problems, and compare them theoretically and empirically. Interesting results include the failure of bidirectional search and close relationships between Dijkstra's algorithm, weighted A*, and depth-first search.

Probabilistic Classifiers Chains (PCC) offers interesting properties to solve multi-label classification tasks due to its ability to estimate the joint probability of the labels. However, PCC presents the major drawback of having a high computational cost in the inference process required to predict new samples. Lately, several approaches have been proposed to overcome this issue, including beam search and an epsilon-Approximate algorithm based on uniform-cost search. Surprisingly, the obvious possibility of using heuristic search has not been considered yet. This paper studies this alternative and proposes an admisible heuristic that, applied in combination with A* algorithm, guarantees, not only optimal predictions in terms of subset 0/1 loss, but also that it always explores less nodes than epsilon-Approximate algorithm. In the experiments reported, the number of nodes explored by our method is less than two times the number of labels for all datasets analyzed. But, the difference in explored nodes must be large enough to compensate the overhead of the heuristic in order to improve prediction time. Thus, our proposal may be a good choice for complex multi-label problems.

The preferences of users are important in route search and planning. For example, when a user plans a trip within a city, their preferences can be expressed as keywords shopping mall, restaurant, and museum, with weights 0.5, 0.4, and 0.1, respectively. The resulting route should best satisfy their weighted preferences. In this paper, we take into account the weighted user preferences in route search, and present a keyword coverage problem, which finds an optimal route from a source location to a target location such that the keyword coverage is optimized and that the budget score satisfies a specified constraint. We prove that this problem is NP-hard. To solve this complex problem, we pro- pose an optimal route search based on an A* variant for which we have defined an admissible heuristic function. The experiments conducted on real-world datasets demonstrate both the efficiency and accu- racy of our proposed algorithms.

Real-time strategy (RTS) games are hard from an AI point of view because they have enormous state spaces, combinatorial branching factors, allow simultaneous and durative actions, and players have very little time to choose actions. For these reasons, standard game tree search methods such as alpha- beta search or Monte Carlo Tree Search (MCTS) are not sufficient by themselves to handle these games. This paper presents an alternative approach called Adversarial Hierarchical Task Network (AHTN) planning that combines ideas from game tree search with HTN planning. We present the basic algorithm, relate it to existing adversarial hierarchical planning methods, and present new extensions for simultaneous and durative actions to handle RTS games.

To engage diverse agents in cooperative behavior, it is important, even necessary, to provide algorithms that do not reveal information that is private or proprietary.A number of recent planning algorithms enable agents to plan together for shared goals without disclosing information about their private state and actions. But these algorithms lack clear and formal privacy guarantees: the fact that they do not require agents to explicitly reveal private information, does not imply that such information cannot be deduced. The main contribution of this paper is an enhanced version of the distributed forward-search planning framework of Nissim and Brafman that reveals less information than the original algorithm, and the first, to our knowledge, discussion and formal proof of privacy guarantees for distributed planning and search algorithms.

Monte-Carlo Tree Search (MCTS) algorithms such as UCT are an attractive online framework for solving planning under uncertainty problems modeled as a Markov Decision Process. However, MCTS search trees are constructed in flat state and action spaces, which can lead to poor policies for large problems. In a separate research thread, domain abstraction techniques compute symmetries to reduce the original MDP. This can lead to significant savings in computation, but these have been predominantly implemented for offline planning. This paper makes two contributions. First, we define the ASAP (Abstraction of State-Action Pairs) framework, which extends and unifies past work on domain abstractions by holistically aggregating both states and state-action pairs -- ASAP uncovers a much larger number of symmetries in a given domain. Second, we propose ASAP-UCT, which implements ASAP-style abstractions within a UCT framework combining strengths of online planning with domain abstractions. Experimental evaluation on several benchmark domains shows up to 26% improvement in the quality of policies obtained over existing algorithms.