John Levine

AUTOMATIC-GENERATION OF TECHNICAL DOCUMENTATION

Natural-language generation (NLG) techniques can be used to automatically produce technical documentation from a domain knowledge base and linguistic and contextual models. We discuss this application of NLG technology from both a technical and a usefulness (costs and benefits) perspective. This discussion is based largely on our experiences with the Intelligent Documentation Advisory System (IDAS) documentation-generation project, and the reactions that various interested people from industry have had to IDAS. We hope that this summary of our experiences with IDAS and the lessons we have learned from it will be beneficial for other researchers who wish to build technical documentation-generation systems.

Automatic Generation of On-Line Documentation in the IDAS Project

The Intelligent Documentation Advisory System generates on-line documentation and help messages from a domain knowledge base, using natural-language (NL) generation techniques. This paper gives an overview of IDAS, with particular emphasis on: (1) its architecture and the types of questions it is capable of answering; (2) its KR and NL generation systems, and lessons we have learned in designing them; and (3) its hypertext-like user interface, and the benefits such an interface brings.

Towards a Video Game Description Language

As participants in this Dagstuhl session address the challenge of General Video Game Playing (GVGP), we have recognised the need to create a Video Game Description Language (VGDL). Unlike General Game Playing, we have envisioned GVGP will not require a prescribed language to facilitate understanding of the logic of the game: requiring the computational agent to ascertain these facts for itself. However, we would still require means to define the wide range of problems the GVGP agents may face for the purpose of classification. Not only would such a language provide means to encapsulate the features and mechanics of a game for the purposes of human understanding, but also provide context for the evaluation of GVGP agents having completed playing. Outside of the issues of classification, there is also the opportunity for automatic game generation. Given the intent of the GVGP group to work within a framework akin to the one of the Physical Travelling Salesman Problem (PTSP), we aim to attach a code-base to the VGDL compiler that derives implementations of these games from the definition that can be used in conjunction with GVGP. Implementing such a compiler could provide numerous opportunities; users could modify existing games very quickly, or have a library of existing implementations defined within the language (e.g. an Asteroids ship or a Mario avatar) that have pre-existing, parameterised behaviours that can be customised for the users specific purposes. Provided the language is fit for purpose, automatic game creation could be explored further through experimentation with machine learning algorithms, furthering research in game creation and design. In order for both of these perceived functions to be realised and to ensure it is suitable for a large user base we recognise that the language carries several key requirements. Not only must it be human-readable, but retain the capability to be both expressive and extensible whilst equally simple as it is general. In our preliminary discussions, we sought to define the key requirements and challenges in constructing a new VGDL that will become part of the GVGP process. From this we have proposed an initial design to the semantics of the language and the components required to define a given game. Furthermore, we applied this approach to represent classic games such as Space Invaders, Lunar Lander and Frogger in an attempt to identify potential problems that may come to light. In summary, our group has agreed on a series of preliminary language components and started to experiment with forms of implementation for both the language and the attached framework. In future we aim to realise the potential of the VGDL for the purposes of Procedural Content Generation, Automatic Game Design and Transfer Learning and how the roadmap for GVGP can provide opportunities for these areas.

General video game playing

One of the grand challenges of AI is to create general intelligence: an agent that can excel at many tasks, not just one. In the area of games, this has given rise to the challenge of General Game Playing (GGP). In GGP, the game (typically a turn-taking board game) is defined declaratively in terms of the logic of the game (what happens when a move is made, how the scoring system works, how the winner is declared, and so on). The AI player then has to work out how to play the game and how to win. In this work, we seek to extend the idea of General Game Playing into the realm of video games, thus forming the area of General Video Game Playing (GVGP). In GVGP, computational agents will be asked to play video games that they have not seen before. At the minimum, the agent will be given the current state of the world and told what actions are applicable. Every game tick the agent will have to decide on its action, and the state will be updated, taking into account the actions of the other agents in the game and the game physics. We envisage running a competition based on GVGP playing, using arcadestyle (e.g. similar to Atari 2600) games as our starting point. These games are rich enough to be a formidable challenge to a GVGP agent, without introducing unnecessary complexity. The competition that we envisage could have a number of tracks, based on the form of the state (frame buffer or object model) and whether or not a forward model of action execution is available. We propose that the existing Physical Travelling Salesman (PTSP) software could be extended for our purposes and that a variety of GVGP games could be created in this framework by AI and Games students and other developers. Beyond this, we envisage the development of a Video Game Description Language (VGDL) as a way of concisely specifying video games. For the competition, we see this as being an interesting challenge in terms of deliberative search, machine learning and transfer of existing knowledge into new domains.

Learning action strategies for planning domains using genetic programming

There are many different approaches to solving planning problems, one of which is the use of domain specific control knowledge to help guide a domain independent search algorithm. This paper presents L2Plan which represents this control knowledge as an ordered set of control rules, called a policy, and learns using genetic programming. The genetic programs crossover and mutation operators are augmented by a simple local search. L2Plan was tested on both the blocks world and briefcase domains. In both domains, L2Plan was able to produce policies that solved all the test problems and which outperformed the hand-coded policies written by the authors.

Investigation of Different Seeding Strategies in a Genetic Planner

Planning is a difficult and fundamental problem of AI. An alternative solution to traditional planning techniques is to apply Genetic Programming. As a program is similar to a plan a Genetic Planner can be constructed that evolves plans to the plan solution. One of the stages of the Genetic Programming algorithm is the initial population seeding stage. We present five alternatives to simple random selection based on simple search. We found that some of these strategies did improve the initial population, and the efficiency of the Genetic Planner over simple random selection of actions.

O-P/sup 3/: Supporting the planning process using open planning process panels

This paper introduces Open Planning Process Panels (O-p3). These panels are based on explicit models of the planning process and axe used to coordinate the development and evaluation of multiple courses of action. We describe the generic ideas behind O-P3 technology, a general methodology for building O-P3 interfaces and two applications based on O-P3 technology the Air Campaign Planning Process Panel (ACP3) and the O-Plan two-user mixedinitiative planning Web demonstration. This work has an impact on a number of important research areas outside planning, including Computer Supported Cooperative Work (CSCW) and workflow support.

Emerging Cooperation With Minimal Effort: Rewarding Over Mimicking

This paper compares supervised and unsupervised learning mechanisms for the emergence of cooperative multiagent spatial coordination using a top-down approach. By observing the global performance of a group of homogeneous agents-supported by a nonglobal knowledge of their environment-we attempt to extract information about the minimum size of the agent neurocontroller and the type of learning mechanism that collectively generate high-performing and robust behaviors with minimal computational effort. Consequently, a methodology for obtaining controllers of minimal size is introduced and a comparative study between supervised and unsupervised learning mechanisms for the generation of successful collective behaviors is presented. We have developed a prototype simulated world for our studies. This case study is primarily a computer games inspired world but its main features are also biologically plausible. The two specific tasks that the agents are tested in are the competing strategies of obstacle-avoidance and target-achievement. We demonstrate that cooperative behavior among agents, which is supported only by limited communication, appears to be necessary for the problems efficient solution and that learning by rewarding the behavior of agent groups constitutes a more efficient and computationally preferred generic approach than supervised learning approaches in such complex multiagent worlds

Improving control through subsumption in the EvoTanks domain

In this paper we further explore the potential of a decentralised controller architecture that places multi-layer perceptrons within a subsumption hierarchy. Previous research exploring this approach proved successful in generating agents that could solve problems while coping with new reactive stimuli. However there were many unresolved questions that we wished to explore. In this paper we explore the use of our architecture with iterative training, increased controller modularity and conflicting goals. Results provide some interesting insights into the potential this method could have to agent designers.

Scaling-up behaviours in EvoTanks: Applying subsumption principles to artificial neural networks

Applying evolution to generate simple agent behaviours has become a successful and heavily used practice. However the notion of scaling up behaviour into something more noteworthy and complex is far from elementary. In this paper we propose a method of combining neuroevolution practices with the subsumption paradigm; in which we generate Artificial Neural Network (ANN) layers ordered in a hierarchy such that high-level controllers can override lower behaviours. To explore this proposal we apply our controllers to the dasiaEvoTankspsila domain; a small, dynamic, adversarial environment. Our results show that once layers are evolved we can generate competent and capable results that can deal with hierarchies of multiple layers. Further analysis of results provides interesting insights into design decisions for such controllers, particularly when compared to the original suggestions for the subsumption paradigm.