Many problems impede the design of multiagent systems, not the least of which is the passing of information between agents. While others hand implement communication routes and semantics, we explore a method by which communication can evolve. In the experiments described here, we model agents as connectionist networks. We supply each agent with a number of communications channels implemented by the addition of both input and output units for each channel. The output units initiate environmental signals whose amplitude decay over distance and are perturbed by environmental noise. An agent does not receive input from other individuals, rather the agent’s input reflects the summation of all other agents’ output signals along that channel. Because we use real-valued activations, the agents communicate using real-valued vectors. Under our evolutionary program, GNARL, the agents coevolve a communication scheme over continuous channels which conveys task-specific information.

The field of adaptive behavior holds that higher-level cognitive skills arise from the more primitive ability of an agent to adapt to its environment. Although many behaviors have been studied in this bottom-up fashion (e.g., obstacle avoidance, wandering, environment exploration, food collection, planning, predator avoidance, locomotion, action selection, flocking, etc.), relatively few people have studied communication as adaptive behavior. In this dissertation, I explore how communication can be understood as an adaptation by agents to their environment. I begin by looking at behavior-based methods for agent evolution, and propose a connectionist version of subsumption which supports learning. I reject this as a method of evolving communication, however, because of certain assumptions the approach requires. I then turn to evolutionary programming, a population-based search technique, and show how it can be used to evolve agents with far fewer assumptions. With this background, I move on to the evolution of communication. I begin with a set of independent agents, instantiated as recurrent neural networks. After arguing against systems which use discrete symbols to evolve communication, I supply my agents with a number of continuous communications channels. The agents use these channels to initiate real-valued signals which propagate through the environment, decaying over distance, perhaps being perturbed by environmental noise. Initially, the agents' signals appear random; over time, a structure emerges as the agents learn to communicate task-specific information about their environment. I demonstrate how different communication schemes can evolve for a task, and then discover a commonality between the schemes in terms of information passed between agents. From this I discuss what it means to communicate, and describe how a semantics emerges in the agents' signals relative to their task domain.

Standard methods for simultaneously inducing the structure and weights of recurrent neural networks limit every task to an assumed class of architectures. Such a simplification is necessary since the interactions between network structure and function are not well understood. Evolutionary computations, which include genetic algorithms and evolutionary programming, are population-based search methods that have shown promise in many similarly complex tasks. This paper argues that genetic algorithms are inappropriate for network acquisition and describes an evolutionary program, called GNARL, that simultaneously acquires both the structure and weights for recurrent networks. GNARL's empirical acquisition method allows for the emergence of complex behaviors and topologies that are potentially excluded by the artificial architectural constraints imposed in standard network induction methods.< >

It has occurred to many researchers to apply genetic algorithms to the training of recurrent neural networks. These studies generally avoid total network induction, i.e., inducing both the topology and parametric values of a network, favoring instead simple parametric learning. Also, they often rely on standard forms of crossover to manipulate network structures, a process which actually inhibits network evolution. In addition, the common commitment to bit string representations introduces an artificial limitation on the range of networks that can be created. In this paper, an evolutionary program, called GNARL, is presented that performs total network induction. GNARL’s ability to induce a range of appropriate solutions is demonstrated on an interesting control task.

From the many possible perspectives in which an agent may be viewed, behavior-based AI selects observable actions as a particularly useful level of description. Yet behavior is clearly not structure, and anyone using behavior-based constraints to construct an agent still faces many implementa-tional roadblocks. Such obstacles are typically avoided by adopting a finite state automaton (FSA) as a base representation. As a result, potential benefits from alternative formalisms are ignored. To explore these benefits, our work adopts a multi-level view of an agent: behaviors and FSAs are but two of many levels of description. We still focus on behaviors for the expression of design constraints, but we avoid using FSAs as an implementation. Our particular agent, Addam, is comprised of a set of connectionist networks, a substrate which promotes the automatic design of subsumptive systems. Moreover, the implementational choice has important behavioral consequences – some complex behaviors emerge due to interactions among networks and need not be specified explicitly. In this way, the underlying layers leak into one another, each affecting the others in subtle and desirable ways.

This paper introduces GNARL, an evolutionary program which induces recurrent neural networks that are structurally unconstrained. In contrast to constructive and destructive algorithms, GNARL employs a population of networks and uses a fitness function's unsupervised feedback to guide search through network space. Annealing is used in generating both gaussian weight changes and structural modifications. Applying GNARL to a complex search and collection task demonstrates that the system is capable of inducing networks with complex internal dynamics.

Cognitive scientists, AI researchers in particular, have long-recognized the enormous benefits of modularity (e.g., Simon, 1969), as well as the need for self-organization (Samuel, 1967) in creating artifacts whose complexity approaches that of human intelligence. And yet these two goals seem almost incompatible, since truly modular systems are usually designed, and systems that truly learn are inherently nonmodular and produce only simple behaviors. Our paper seeks to remedy this shortcoming by developing a new architecture of Additive Adaptive Modules which we instantiate as Addam, a modular agent whose behavioral repertoire evolves as the complexity of the environment is increased.1