On the Reactive Limits of Autonomous Agents zyx George Fodor ABB Industrial Systems AB, ISY/AMC S-721 67 Vasteras, Sweden and Department of Electrical and Computer Westeni Michigan University, Kalamazoo, Engineering MI 49008-5066, USA Abstract zyxwvutsrq - Autonomous agents capable to perform unattended complex sequential tasks, are often required to detect and recover from unexpected changes in their environment. In this paper, the recovery limits of an autonomous agent with a given reactive machinery are considered with respect to the agent’s capability to detect the existence of unexpected disturbances. I. INTRODUCTION Autonomous agents capable to perform unattended complex sequential tasks, are often required to detect and recover from unexpected changes in their environment. The problem of determining the answcr by zyxwvuts an autonomous agent to unexpected changes in its environment has been notoriously difficult. A more confined question Uiat can be examined is ahout Uic limits of an autonomous agent with a given reactive rnachinery to recover from an uncxpcctcd disturbance in its environment. These limits are related to the agent’s capability to detect those disturbances. The main characteristics of autonomous agents considered in this research are shown in Section 2. The identification operation described in Scctioii 3 is based upon the tlicory of ontological control [3], zyxwvutsrq [5]. The recovery operation is described in Section 4 and it is based on the fuzzy automata theory of Grantrier arid Patyra I(,], [lo[. The results of current research indicate that an autonomous agent which has high capability to detect that an unexpected change has occurred in its environment, has in turn limited capability to give a proper recovery answer to that change. Alternatively, an agent that has a large range of Janos L. Grantner Engineering Department of Electrical and Computer Western Michigan University, Kalamazoo, MI 49008-5066, USA answers, that are devised from a real-time re- planning level, can have only a limited capability to detect Uiat such a change has occurred. 11. AUTONOMOUS AGENTS Many unattended applications in tlic areas of robotics, marine vessels and avionics need autonomous systems capable to act consistently with respect to the initial goals when they encounter unexpected situations in their environment. Such systems comprise of a reactive level that interacts directly with their environment through scnsors and actuators and which can perform control tasks such as posi lion i ng, or scq ueii ci ti g . A super visor y lcvel of I ~ C syslcni cart detect wlien llic rcaclive lcvcl cannot cope with, or does not intend to continue with the current control task, and using additional stored knowledge, will perform a recovery operation, or a re-planning of the rcactivc control lcvel. The control algorilhm of the reactive level is assumed to have tlie following properties: I. The algorithni is sequential und has a zyx state- dependent represenlitriorz. The section of the algorithm which determines the agent’s reactive answers to known situations is sequential and is state-based. Continuous control sections are subsumed into the sequential state representation by relevant domain intervals in whicli the continuous control has a pre-defined stable behavior, or quality. 2. Goal states. The goal states are explicitly represented, and the sequences of states leading to goal states are either ’explicitly or zy 427