The standard approach to decision tree induction is a top-down, greedy algorithm that makes locally optimal, irrevocable decisions at each node of a tree. In this paper, we study an alternative approach, in which the algorithms use limited lookahead to decide what test to use at a node. We systematically compare, using a very large number of decision trees, the quality of decision trees induced by the greedy approach to that of trees induced using lookahead. The main results of our experiments are: (i) the greedy approach produces trees that are just as accurate as trees produced with the much more expensive lookahead step; and (ii) decision tree induction exhibits pathology, in the sense that lookahead can produce trees that are both larger and less accurate than trees produced without it. 1. Introduction The standard algorithm for constructing decision trees from a set of examples is greedy induction --- a tree is induced top-down with locally optimal choices made at each node, with...

Abstract. Years of work have gone into algorithms and optimizations for twoplayer perfect-information games such as Chess and Checkers. It is only more recently that serious research has gone into games with imperfect information, such as Bridge, or game with more than two players or teams of players, such as Poker. This work focuses on multi-player game search in the card games Hearts and Spades, providing an overview of past research in multi-player game search and then presents new research results regarding the optimality of current search techniques and the need for good opponent modeling in multi-player game search. We show that we are already achieving near-optimal pruning in the games Hearts and Spades. 1

The standard approach t decision tree in duction is a top-down greedy agonthm that makes locall} optimal irrevocable decisions at each node of a tree In this paper we empir-•call} study an alternative approach in which the algorithms use one-level loo kalie to deride what test to use at a node weystematically compare using a very large number of rfal and artificial data sets the quality of dmsion trees induced by the greedv approach to that of trees induced using lookahead The main observations from our experments are (1) the greedv approach consistently produced trees that were just as at curate as trees produced with the much more expensive lookahead step and (n) we observed manv instances of pathology, le, lookalnad producrd trees that were both larger and less accurate than trees produced without it 1

Game-tree search algorithms, such as the two-player Minimax algorithm and its multi-player counterpart, Max-n, are a fundamental component in the development of computer programs for playing extensive-form games. The success of these algorithms is limited by the underlying assumptions on which they are built. For example, it is traditionally assumed that deeper search always produces better decisions and also that search procedures can assume all players are selfish and ignore social orientations. Deviations from these assumptions can occur in real games and can affect the success of a traditional search algorithms. The goal of my thesis is to determine when such deviations occur and modify the search procedure to correct the errors that are introduced.