Christian J. Darken

Fast learning in networks of locally-tuned processing units

We propose a network architecture which uses a single internal layer of locally-tuned processing units to learn both classification tasks and real-valued function approximations (Moody and Darken 1988). We consider training such networks in a completely supervised manner, but abandon this approach in favor of a more computationally efficient hybrid learning method which combines self-organized and supervised learning. Our networks learn faster than backpropagation for two reasons: the local representations ensure that only a few units respond to any given input, thus reducing computational overhead, and the hybrid learning rules are linear rather than nonlinear, thus leading to faster convergence. Unlike many existing methods for data analysis, our network architecture and learning rules are truly adaptive and are thus appropriate for real-time use.

Fast adaptive k -means clustering: some empirical results

The authors present learning rate schedules for fast adaptive k-means clustering which surpass the standard MacQueen learning rate schedule (J. MacQeen, 1967) in speed and quality of solution by several orders of magnitude for large k. The methods accomplish this by largely overcoming the problems of metastable local minima and nonstationarity of cluster region boundaries which plague the MacQueen approach. The authors use simulation results to compare the clustering performances of four learning rate schedules applied to independently sampled data from a uniform distribution in one and two dimensions

Learning rate schedules for faster stochastic gradient search

The authors propose a new methodology for creating the first automatically adapting learning rates that achieve the optimal rate of convergence for stochastic gradient descent. Empirical tests agree with theoretical expectations that drift can be used to determine whether the crucial parameter c is large enough. Using this statistic, it will be possible to produce the first adaptive learning rates which converge at optimal speed.<<ETX>>

A computational model for human eye-movements in military simulations

Models of eye movements of an observer searching for human targets are helpful in developing accurate models of target acquisition times and false positive detections. We develop a new model describing the distribution of gaze positions for an observer which includes both bottom-up (salience) and top-down (task dependent) factors. We validate the combined model against a bottom-up model from the literature and against the bottom up and top down parts alone using human performance data on stationary targets. The new model is shown to be significantly better. The new model requires a large amount of data about the terrain and the target that is obtained directly from the 3D simulation through an automated process.

Modeling and integration of situational awareness and soldier target search

Representation of search and target acquisition (STA) in military models and simulations arguably abstracts the most critical aspects of combat. This research focuses on the search aspect of STA for the unaided human eye. It is intuitive that an individual’s environmental characteristics and interpretation of the environment in the context of all comprehended information, commonly summarized as their situational awareness (SA), influences attention and search. Current simulation models use a primitive sweeping search method that devotes an unbiased amount of time to every area in an entity’s field of regard and neglects the effects of SA. The goal of this research is to provide empirical results and recommend modeling approaches that improve the representation of unaided search in military models and simulations. The major contributions towards this goal include novel empirical results from two incremental eye-tracking experiments, analysis and modeling of the eye-tracking data to illustrate the effect of the environment and SA on search, and a recommended model for unaided search for high-fidelity combat simulation models. The results of this work support soldier search models driven by metrics that summarize the threat based on environmental characteristics and contextual information.

Computer Graphics-Based Models Of Target Detection: Algorithms, Comparison To Human Performance, And Failure Modes

Target detection is one of the fundamental phenomena that must be modeled in military simulations. When the target detection model fails, entities that should not be mutually aware engage, and entities that should fight ignore one another. The potential negative consequences for training and analysis are obvious. We describe three closely related computer graphics-based detection models for virtual simulation that can avoid some of the limitations of previous approaches. These models incorporate a standard target detection model, but feed it with more accurate target exposure and contrast data than has been done previously. Two variants of the base model attempt to improve the target contrast calculation and add color sensitivity. We compare the predictions of these models to human performance to show that the model variants have their intended effect. The performance of even the best models can deviate drastically from the performance of the human eye under some circumstances represented in our experiment. We lump these into categories as an aid to understanding the state of the art and to motivate future research.

Behavior Selection Using Utility-Based Reinforcement Learning in Irregular Warfare Simulation Models

The Theory of Planned Behavior TPB provides a conceptual model for use in assessing behavioral intentions of humans. Agent based social simulations seek to represent the behavior of individuals in societies in order to understand the impact of a variety of interventions on the population in a given area. Previous work has described the implementation of the TPB in agent based social simulation using Bayesian networks. This paper describes the implementation of the TPB using novel learning techniques related to reinforcement learning. This paper provides case study results from an agent based simulation for behavior related to commodity consumption. Initial results demonstrate behavior more closely related to observable human behavior. This work contributes to the body of knowledge on adaptive learning behavior in agent based simulations.

Representing trust in cognitive social simulations

Trust plays a critical role in communications, strength of relationships, and information processing at the individual and group level. Cognitive social simulations show promise in providing an experimental platform for the examination of social phenomena such as trust formation. This paper describes the initial attempts at representation of trust in a cognitive social simulation using reinforcement learning algorithms centered around a cooperative Public Commodity game within a dynamic social network.

A practical situation based agent architecture for social simulations

The concept of situation is central to the decision making processes of both human and software agents. The recognition of situation facilitates decision processes that ultimately result in action selection. Cognitive agent architectures that incorporate the concept of situation provide the opportunity for more sophisticated representations of human behavior and for more sophisticated decision support applications. This paper provides an overview of a general cognitive architecture for use in multi-agent simulation with the concept of situation central to the action selection and decision making process.

Learning and prediction of relational time series

Abstract Learning to predict events in the near future is fundamental to human and artificial agents. Many prediction techniques are unable to learn and predict a stream of relational data online when the environments are unknown, non-stationary, and no prior training examples are available. This paper addresses the online prediction problem by introducing a low complexity learning technique called Situation Learning and several prediction techniques that use the information from Situation Learning to predict the next likely event. The prediction techniques include two variants of a Bayesian inference technique, a variable order Markov model prediction technique and situation matching techniques. We compared their prediction accuracies quantitatively for three domains: a role-playing game, computer network intrusion system alerts, and event prediction of maritime paths in a discrete-event simulator.