This thesis documents an investigation into the automatic provision of reasonably realistic motion for the computer generated target ships in a marine simulator. The thesis explores: automatic collision avoidance between the target ships; automatic track keeping for the target ships; the use of sea stabilised and land stabilised motion for the target ships; some issues of software fault tolerance in marine simulators; message frameworks for use in a Multi-Agent System (MAS) simulation; the opportunity to provide different manoeuvring characteristics for different target ships; and the use of autonomous agents to control the target ships. A software system has been developed to facilitate this research. Entitled "A Multi-Agent Realm for Investigating Navigators' Educational Simulators" (MARINES), the software is a MAS providing much of the functionality of a marine simulator instructor station; basic functions are encapsulated into the instructor environment and additional features are provided by processes that connect to the environment using Dynamic Data Exchange. The processes can also connect to each other and, in MARINES, co-operate to navigate the ships. These co-operative, autonomous processes are the agents that together form a MAS. A simple 3D view is also connected, enabling the view from the bridge of a specific target ship to be assessed. The MARINES software is written using C++ to run under Microsoft Windows v3.1. Therefore, the processes multi-task co-operatively. In MARINES each target ship can be made to perform in an individual manner; manoeuvring and performance characteristics can be customised to simulate a specific ship type. Additionally, the agents performing collision avoidance can be given rule sets that interpret the International Regulations for the Prevention of Collisions at Sea in subtly different ways, and the track-keeping agents can have different beliefs about the manoeuvring capabilities of the vessels they control. Automatic collision avoidance and track keeping is performed for two-ship situations even when the set and drift of a current is introduced. A comparison is made with the tracks of land stabilised targets. This shows how aspect, course and speed are affected by ignoring the effects of the current, and demonstrates the need for an accurate simulation.

Large-scale processes that are modeled using differential algebraic equations based on mass and energy balance calculations at times require excessive computation time to simulate. Depending on the complexity of the model, these simulations may require many iterations to converge and in some cases they may not converge at all. Application of a storage and retrieval technique, named in situ adaptive tabulation or ISAT is proposed for faster convergence of process simulation models. Comparison with neural networks is performed, and better performance using ISAT for extrapolation is shown. In particular, the requirement of real-time dynamic simulation is discussed for operating training simulators (OTS). Integration of ISAT to a process simulator (CHEMCAD®) using the input-output data only is shown. A regression technique based on partial least squares (PLS) is suggested to approximate the sensitivity without accessing the first-principles model. Different record distribution strategies to build an ISAT database are proposed and better performance using the suggested techniques is shown for different case studies. A modified ISAT algorithm (mISAT) is described to improve the retrieval rate, and its performance is compared with the original approach in a case study. State estimation is a key requirement of many process control and monitoring strategies. Different nonlinear state estimation techniques studied in the past are discussed with their relative advantages/disadvantages. A robust state estimation technique like moving horizon estimation (MHE) has a trade-off between accuracy of state estimates and the computational cost. Implementation of MHE based ISAT is shown for faster state estimation, with an accuracy same as that of MHE. Flowsheet optimization aims to optimize an objective or cost function by changing various independent process variables, subject to design and model constraints. Depending on the nonlinearity of the process units, an optimization routine can make a number of calls for flowsheet (simulation) convergence, thereby making the computation time prohibitive. Storage and retrieval of the simulation trajectories can speed-up process optimization, which is shown using a CHEMCAD® flowsheet. Online integration of an ISAT database to solve the simulation problem along with an outer-loop consisting of the optimization routine is shown using the sequential-modular approach. / text

Simulation-based training has gained widespread attention recently as a response to drawbacks associated with traditional training approaches, such as high training costs (instructors, equipment, etc.), high risks (e.g. pilot training), and ethical issues (e.g. medical training), as well as a lack of availability of certain training environments (e.g. space exploration). Apart from their target training domains, many of aspects of simulations differ, such as their degree of physical realism (fidelity), scenarios (e.g. story), and pedagogical aspects (e.g. after-action reviews and collaborative learning). Among those aspects, designers have mostly focused on developing high-fidelity simulations with the expectation of increasing the effectiveness of training. However, some authors suggest that the above belief is a myth as researchers have failed to identify a linear relationship between the (physical) fidelity and training effectiveness of simulations.  Most researchers have therefore evaluated the correspondence between the behaviours of trainees in both real world and simulated contexts, however, the existing methods of simulation validation using behavioural measures have a number of drawbacks, such as the fact that they do not address certain complex phenomena of skills acquisition. Bridging the above knowledge gap, this research reports on empirical investigations using an improved methodology for validating training simulations. This research includes an investigation of the user experience of trainees, with respect to the acceptance of virtual scenarios provoking a similar psychophysiological response as in real world scenarios, and the training potential of simulations with respect to the positive transfer of training from a simulator to real world operational contexts. The most prominent features of the proposed methodology include the use of psychophysiological measures in addition to traditional behavioural measures and the use of natural (quasi-) experiments. Moreover, its conceptual framework was influenced by contemporary theories in cognitive science (e.g. constructivism and embodied cognition). The results of this research have several important theoretical and methodological implications, involving, for example, the dependency of the effectiveness of simulations on the perceived realism of trainees, which is more embodied than has been predicted by previous researchers, and the requirement of several different types/levels of adaptive training experience, depending on the type of trainee. / Träning i simulatorer har på senare år fått ökad uppmärksamhet som en respons på problem och svårigheter förknippade med traditionella träningsansatser, såsom höga kostnader (instruktörer och utrustning, etc.), hög risk (t.ex. träning av piloter), och etiska aspekter (t.ex. träning av kirurger), likaväl som avsaknaden av träningsmöjligheter och miljöer (t.ex. forskning om rymden). Bortsett från vad som specifikt tränas så skiljer sig simuleringar åt i ett flertal olika aspekter såsom fysisk realism (eng. fidelity), scenarier (handling) och pedagogiska aspekter (t.ex. genomgång efter övning och kollaborativt lärande).  Bland dessa aspekter så har designers ofta fokuserat att utveckla simuleringar med hög realism med förväntningen att detta ska göra träningen mer effektiv. Litteraturen antyder dock att denna föreställning inte stämmer och att de flesta simuleringar med hög realism inte har lyckats uppnå denna målsättning. En slutsats är därför att det finns ett behov av metoder som kan validera potentialen hos simuleringar avsedda att stödja träning – redan innan dessa används. Enligt litteraturen så är utbildningspotentialen hos en simulering starkt kopplad till hur väl den psykologiska effekten en simulering har, stämmer överens med en verklig upplevelse. Forskning har emellertid identifierat ett flertal svagheter hos existerande ansatser för att validera simuleringar; de är oftast baserade på prestations- och/eller subjektiva mätningar; de har fokuserat en eller ett fåtal psykologiska aspekter; och de bygger på traditionella teorier. Baserat på resultat från studier av en kör-simulator presenteras och föreslås i denna avhandling ett förbättrat ramverk för utvärdering. De mest centrala egenskaperna hos det föreslagna ramverket inbegriper användandet av psyko-fysiologiska mått tillsammans med mer traditionella mått; det konceptuella ramverket bygger på samtida teoretiska ansatser (tex konstruktivism och kroppslig kognition); samt användandet av fält (kvasi-) experiment. Utöver uppnåendet av uppsatta mål för forskningen så har resultaten ett flertal teoretiska och metodologiska implikationer. Bland dessa återfinns beroendet mellan effektiviteten hos en simulering och den upplevelse av realitet som de tränade har, vilken är mer grundläggande än vad som rapporterats i tidigare forskning, samt kravet på flera och olika typer av anpassning av träningsupplevelse för den tränade för att förhöja potentialen hos träningssimulatorer. / SIDA Funded National e-Learning Centre Project at the University of Colombo School of Computing, Sri Lanka

Behavioral animation has become popular for creating virtual characters that are autonomous agents and thus self-animating. This is useful for lessening the workload of human animators, populating virtual environments with interactive agents, etc. Unfortunately, current behavioral animation techniques suffer from three key problems: (1) deliberative behavioral models (i.e., cognitive models) are slow to execute; (2) interactive virtual characters cannot adapt online due to interaction with a human user; (3) programming of behavioral models is a difficult and time-intensive process. This dissertation presents a collection of papers that seek to overcome each of these problems. Specifically, these issues are alleviated through novel machine learning schemes. Problem 1 is addressed by using fast regression techniques to quickly approximate a cognitive model. Problem 2 is addressed by a novel multi-level technique composed of custom machine learning methods to gather salient knowledge with which to guide decision making. Finally, Problem 3 is addressed through programming-by-demonstration, allowing a non technical user to quickly and intuitively specify agent behavior.