Towards a Generic Infrastructure to Adjust the Autonomy of Soar Agents

Developing and testing intelligent agents is a complex task that is both time-consuming and costly. This creates the potential that problems in the agent’s behavior will be realized only after the agent has been put to use. In this paper we explore two implementations of a generic agent selfassessment framework applied to the Soar agent architecture. Our system extends previous work and can be used to achieve adjustable levels of agent autonomy or runtime verification with only minor modifications to existing Soar agents. We present results indicating the computational overhead of both approaches compared against an agent that exhibits identical behavior without the help of the self-assessment framework. Agents whose behavior has not been completely validated run the risk of performing their tasks incorrectly. Such situations may occur if an agent encounters situations it was not designed to deal with or if its knowledge for how to deal with a particular situation is incorrect. In the former case, the agent is operating outside of its intended (specified) scope; in the latter case, the agent’s implementation is inconsistent with its specification and thus incorrect. Regardless, the impact of unintended behavior may be relatively minor, or, in mission critical situations such as when controlling an unmanned aerial vehicle, the consequences may have a far reaching impact. Prior work on adjustable autonomy (e.g., [Bradshaw et al.2004, Scerri et al.2003, Sellner et al.2006]) has explored a variety of approaches that allow a human supervisor, or an agent itself to deal with exceptional situations that may arise due to incomplete or incorrect knowledge. In this paper, we briefly describe a generic infrastructure of adjustable autonomy that can be used in conjunction with Soar agents and compare the performance of two implementations of this infrastructure. Our generic framework (see Figure 1) works in conjunction with the Soar [Laird, Newell, & Rosenbloom1987] agent architecture. After a Soar agent identifies possible operators (goals or actions) to pursue, the framework intercepts decision making and compares the actions and goals to an external policy. The result of the comparison may deny, require or permit operators the agent has proposed. If all proposed operators have been denied, the framework provides a Copyright c © 2008, Association for the Advancement of Artificial Intelligence (www.aaai.org). All rights reserved. Agent’s Internal Reasoning