<p> Real-Time Strategy games have become a new frontier of artificial intelligence research. Advances in real-time strategy game AI, like with chess and checkers before, will significantly advance the state of the art in AI research. This thesis aims to investigate using heuristic search algorithms to generate effective micro behaviors in combat scenarios for real-time strategy games. <i> Macro</i> and <i>micro</i> management are two key aspects of real-time strategy games. While good macro helps a player collect more resources and build more units, good micro helps a player win skirmishes against equal numbers of opponent units or win even when outnumbered. In this research, we use influence maps and potential fields as a basis representation to evolve micro behaviors. We first compare genetic algorithms against two types of hill climbers for generating competitive unit micro management. Second, we investigated the use of case-injected genetic algorithms to quickly and reliably generate high quality micro behaviors. Then we compactly encoded micro behaviors including influence maps, potential fields, and reactive control into fourteen parameters and used genetic algorithms to search for a complete micro bot, <i> ECSLBot.</i> We compare the performance of our ECSLBot with two state of the art bots, <i>UAlbertaBot</i> and <i>Nova,</i> on several skirmish scenarios in a popular real-time strategy game <i>StarCraft. </i> The results show that the ECSLBot tuned by genetic algorithms outperforms UAlbertaBot and Nova in kiting efficiency, target selection, and fleeing. In addition, the same approach works to create competitive micro behaviors in another game <i>SeaCraft.</i> Using parallelized genetic algorithms to evolve parameters in SeaCraft we are able to speed up the evolutionary process from twenty one hours to nine minutes. We believe this work provides evidence that genetic algorithms and our representation may be a viable approach to creating effective micro behaviors for winning skirmishes in real-time strategy games.</p>

No description available.

Weaver is a hybrid knowledge discovery environment which fills a current gap in Artificial Intelligence (AI) applications, namely tools designed for the development and exploration of existing knowledge in <I>complex, knowledge and data-poor domains. </I>Such domains are typified by incomplete and conflicting knowledge, and data which are very hard to collect. Without the support of robust domain theory, many experimental and modelling assumptions have to be made whose impact on field work and model design are uncertain or simply unknown. Compositional modelling, experimental simulation, inductive learning, and experimental reformulation tools are integrated within a methodology analogous to Popper's scientific method of <I>critical discussion. </I>The purpose of Weaver is to provide a 'laboratory' environment in which a scientist can develop domain theory through an iterative process of <I>in silico</I> experimentation, theory proposal, criticism, and theory refinement. After refinement within Weaver, this domain theory may be used to guide field work and model design. Weaver is a pragmatic response to tool development in complex, knowledge- and data- poor domains. In the compositional modelling tool, a domain-independent algorithm for <I>dynamic multiple scale bridging </I>has been developed. The multiple perspective simulation tool provides an object class library for the construction of multiple simulations that can be flexibly and easily altered. The experimental reformulator uses a simple domain-independent heuristic search to help guide the scientist in selecting the experimental simulations that need to be carried out in order to critically test and refine the domain theory. An example of Weaver's use in an ecological domain is provided in the exploration of the possible causes of population cycles in red grouse (<I>Lagopus, lagopus scoticus</I>). The problem of AI tool validation in complex, knowledge- and data-poor domains is also discussed.

This thesis presents the RAP system: a Reasoner About Physiology. RAP consists of two modules: knowledge representation and reasoning. The knowledge representation module describes commonsense anatomy and physiology at various levels of abstraction and detail. This representation is broad (covers several physiological systems), dense (the number of relationships between anatomical and physiological elements is high) and uniform (the same kind of formalism is used to represent anatomy, physiology and their interrelationships). These features lead to a 'natural' representation of naive physiology which is, therefore, easy to understand and use. The reasoning module performs two tasks: 1) it infers the behaviour of a complex physiological process using the behaviours of its subprocesses and the relationships between them; 2) it reasons about the effect of introducing a fault in the model. In order to reason about the behaviour of a complex process, RAP uses a mechanism which consists of the following tasks: (i) understanding how subprocesses behave; (ii) comprehending how these subprocesses affect each others behaviours; (iii) "aggregating" these behaviours together to obtain the behaviour of the top level process; (iv) giving that process a temporal context in which to act. RAP uses limited commonsense knowledge about faults to reason about the effect of a fault in the model. It discovers new processes which originate as a consequence of a fault and detects processes which misbehave due to a fault. The effects of both newly generated and misbehaving processes are then propagated throughout the model to obtain the overall effect of the fault. RAP represents and reasons about naive physiology and is a step forward in the development of systems which use commonsense knowledge.

This thesis describes a formal analysis of nonmonotonic inheritance. The need for such an understanding of inheritance has been apparent from the time that multiple inheritance and exceptions were mixed in the same representation with the result that the meaning of an inheritance network was no longer clear. Many attempts to deal with the problems associated with nonmonotonic multiple inheritance appeared in the literature but, probably due to the lack of clear semantics there was no general agreement on how many of the standard examples should be handled. This thesis attempts to resolve these problems by presenting a framework for a family of path based inheritance reasoners which allows the consequences of design decisions to be explored. Many of the major theorems are therefore proved without the need to make any commitment as to how conflicts between nonmonotonic chains of reasoning are to be resolved. In particular it is shown that consistent sets of conclusions, known as expansions, exist for a wide class of networks. When commitment is made to a method of choosing between conflicting arguments, particular inheritance systems are produced. The systems described in this thesis can be divided into three classes. The simplest of these, in which an arbitrary choice is made between conflicting arguments, is shown to be very closely related to default logic. The other classes each of which contain four systems, are the decoupled and coupled inheritance systems which use specificity as a guide to choosing between conflicting arguments. In a decoupled system the results relating to a particular node are not affected in any way by derived results concerning other nodes in the inheritance network, whereas in a coupled system decisions in the face of ambiguity are linked to produce expansions which are more intuitively acceptable as a consistent view of the world. A number of results concerning the relationship between these systems are given. In particular it is shown that the process of coupling will not affect the results which lie in the intersection of the expansions produced for a given network.

No description available.

No description available.

No description available.

No description available.

<p> Action selection for an autonomous agent was studied within the confines of truck task scheduling. An experimental setup was established to compare a naive selection approach, a simple market-based optimization approach, and a learning-assisted market-based optimization over a series of scenarios with varying complexity. For sufficiently complex scenarios, the results showed that learning was able to improve the performance of the truck by delaying delivery to a given site until it was the most protable action available. This research adds to the existing autonomous planning research by demonstrating a novel approach for planning under resource constraints. This approach improves upon an existing market-based optimization technique through the use of on-line reinforcement learning for market adjustment.</p>