In this paper, we introduce a case-based method of initializing genetic algorithms that are used to guide search in changing environments. This is incorporated in an anytime learning system. Anytime learning is a general approach to continuous learning in a changing environment. The agent's learning module continuously tests new strategies against a simulation model of the task environment, and dynamically updates the knowledge base used by the agent on the basis of the results. The execution module includes a monitor that can dynamically modify the simulation model based on its observations of the external environment; an update to the simulation model causes the learning system to restart learning. Previous work has shown that genetic algorithms provide an appropriate search mechanism for anytime learning. This paper extends the approach by including strategies, which are learned under similar environmental conditions, in the initial population of the genetic algorithm. Experiments s...

This paper summarizes recent research on competition-based learning procedures performed by the Navy Center for Applied Research in Artificial Intelligence at the Naval Research Laboratory. We have focused on a particularly interesting class of competition-based techniques called genetic algorithms. Genetic algorithms are adaptive search algorithms based on principles derived from the mechanisms of biological evolution. Recent results on the analysis of the implicit parallelism of alternative selection algorithms are summarized, along with an analysis of alternative crossover operators. Applications of these results in practical learning systems for sequential decision problems and for concept classification are also presented. INTRODUCTION One approach to the design of more flexible computer systems is to extract heuristics from existing adaptive systems. We have focused on a class of learning systems that use competition-based procedures, called genetic algorithms (GAs). GAs are ba...

Genetic Algorithms are used to learn navigation and collision avoidance behaviors for robots. The learning is performed under simulation, and the resulting behaviors are then used to control the actual robot. THE LEARNING PARADIGM The approach to learning behaviors for robots described here reflects a particular methodology for learning via a simulation model. The motivation is that making mistakes on real systems may be costly or dangerous. In addition, time constraints might limit the number of experiences during learning in the real world, while in many cases, the simulation model can be made to run faster than real time. Since learning may require experimenting with behaviors that might occasionally produce unacceptable results if applied to the real world, or might require too much time in the real environment, we assume that hypothetical behaviors will be evaluated in a simulation model (the off-line system). As illustrated in Figure 1, the current best behavior can be placed in...

This chapter describes a general framework for multistrategy learning. One idea of this framework is to view learning as an inference process and to integrate the elementary inferences that are employed by the single-strategy learning methods. Another idea is to base learning on building and generalizing a special type of explanation structure called plausible justification tree which is composed of different types of inference and relates the learner's knowledge to the input. In this framework, learning consists of extending and/or improving the knowledge base of the system so that to explain the input received from an external source of information. The framework is illustrated with a specific method that integrates deeply and dynamically explanation-based learning, determination-based analogy, empirical induction, constructive induction, and abduction. 1

This is a progress report on our efforts to design intelligent robots for complex environments. The sort of applications we have in mind include sentry robots, autonomous delivery vehicles, undersea surveillance vehicles, and automated warehouse robots. We are investigating the issues relating to machine learning, using multiple mobile robots to perform tasks such as playing hide-and-seek, tag, or competing to find hidden objects. We propose that the knowledge acquisition task for autonomous robots be viewed as a cooperative effort between the robot designers and the robot itself. The robot should have access to the best model of its world that the designer can reasonably provide. On the other hand, some aspects of the environment will be unknown in advance. For such aspects, the robot itself is in the best position to acquire the knowledge of what to expect in its world. We have implemented these ideas in an arrangement we call case-based anytime learning. This system starts with a pa...

. Machine learning offers the possibility of designing intelligent systems that re fine and improve their initial knowledge through their own experience. This article focuses on the problem of learning sequential decision rules for multi-agent environments. We describe the SAMUEL learning system that uses genetic algorithms and other competition based techniques to learn decision strategies for autonomous agents. One of the main themes in this research is that the learning system should be able to take advantage of existing knowledge where available. This article describes some of the mechanisms for expressing existing knowledge in SAMUEL, and explores some of the issues in selecting constraints for the learning system. 1 Introduction Machine learning offers the possibility of designing intelligent systems through a cooperative effort between humans and machines. The human designers might provide initial guidance to the system in the form of advice obtained through traditional knowled...

SAMUEL is a machine learning program that uses genetic algorithms and other competition-based heuristics to solve sequential decision problems. The system actively explores the space of alternative decision policies in simulation, and modifies its candidate policies based on this experience. Policies are represented as condition-action rules. The genetic algorithm in SAMUEL includes the standard methods of fitness-directed reproduction of rulebases, random mutation, and crossover. In addition, SAMUEL features several Lamarckian operators that modify decision rules on the basis of observed interaction with the task environment. SAMUEL has been used to learn decision policies for behaviors such as navigation and collision avoidance, tracking, and herding, for robots and other autonomous vehicles. SAMUEL also includes mechanisms to allow coevolution of multiple behaviors simultaneously. SAMUEL incorporates a convenient language for the decision rules, making it possible for the user to in...