Interaction, Innovation, and Immunity: Enabling Agents to Play

摘要

In a dynamic environment, an autonomous system has to interact with unknowns in order to remain viable. This interaction exposes the system to problems that need solutions not derivable entirely from the system itself. The autonomous system then has to take somewhat of a leap into the unknown without breaking down. After reviewing basic concepts of play and self, I will model this situation as a process of self-repair in confronting unpredictable and previously unknowable situations. The capacity of an autonomous system to innovate appears then as an immune mechanism for interacting in a dynamic environment. Nature is a rich source of models for constructing artificial agents, from robots to software agents. In this paper I will focus on one neglected natural feature which can enhance the autonomy of agents: that is the capacity of living beings to play. Play is a basic component that helps organisms prosper in dynamic environments. They are able to wiggle about rather safely and in the process they can develop ways to deal with changing situations that could otherwise have fatal consequences. It is as if play were a crucial part of the immune system of an organism. Without play, life may not have the capacity to evolve and adapt to new situations. It is important then to understand how to model effectively this activity. The capacity to play should help constructed agents to operate in changing, uncertain environments. Imagine an agent facing a new situation that extends beyond whatever choices it can make with clear outcomes. In other words, the agent faces uncertainty. How would it proceed on its own? One simple solution is to make choices randomly. But each choice will have to be tested to see how it works and this could become a staggering task for even simple situations. Chances are that many outcomes will be fatal. How could an agent reduce the number of tests and survive to continue functioning autonomously in the midst of uncertainty? In this paper I propose that to enhance its autonomy in an uncertain environment, an agent has to follow nature's example and be able to interact playfully. This in turn allows the agent to learn and modify future interactions. Learning through play augments the agent's capability to develop innovative solutions to problematic situations. In this way, the agent increases its immunity to damages from interactions with the environment. This raises a final issue: learning is not a neutral, passive activity. It is selective and focused. Otherwise it quickly yields an overwhelming amount of raw information. To deal with this problem, an agent has to be able to select and process only what it wants. This sense of desire depends on the nature of the agent. It hinges on what we could picture as its sense of self. The issue becomes then how to design a self-aware agent, so to speak? Would this playful design strengthen a constructed agent's autonomy in uncertain environments? Can play form a loop of behaviors that helps the agent innovate, boost its immunity, and go on interacting? To examine this question I will discuss the following topics: • What is play? • Where does an agent's sense of self come from? • Why does play enhance an agent's autonomy in an uncertain environment? • How do we build play into an agent's behavior?