A Cognitive Model of Situated Autonomy

We introduce situated autonomy and present it as part of the process of action selection. We then discuss the cognitive ingredients of situated autonomy and derive a degree of situated autonomy. 1 Situated Autonomy and Action Selection Autonomous agents have been defined to be agents with self-generated goals using the agent’s motivations [4]. Such agents perform action selection, which is the function of selecting the most relevant and meaningful action [9], entirely for selfish reasons [7]. We believe circumstances of everyday agents provide opportunities for reasoning about relative levels of autonomy. Instead of being autonomous in the general sense, we will focus on the notion of autonomy in the context of a situation and in a team with other agents. We will consider agents able to perform autonomy considerations very fast while they are in the situation. Imagine in a game of basketball, the agent who is in the midst of running anticipates a block and reflects about whether to pass the ball or to run with it. Here, autonomy is a split-second situated assessment. The player considers choosing each of the behaviors “pass the ball” and “run with the ball.” The agent’s considerations of autonomy involve the higher-level goals of scoring or driving the ball into the opponent zone. The agent decides between its orientation to “pass the ball” which means sharing its autonomy toward scoring/driving with another player or its orientation to “run with the ball” which means self-autonomy. Situatedness is to consider situations in the environment as integral component of the agent’s process of deliberation or reactive response generation. Situation is a state of the world as it pertains to a problem. We define situated autonomy as an agent’s stance, as well as the cognitive function of forming the stance, toward assignment of the performer of a goal at a particular moment when facing a particular situation. Assumption of individual versus social rationality affects the cognitive function. At a coarse level the agent’s orientation toward the goal will be whether to abandon it or to decide its overall position toward the goal: to make it an entirely personal goal, to make a goal for another agent, to consider the goal a collaborative effort, or to consider an inclination for the goal that is less than totally self-directed. Here, we are not concerned about responsibility for a goal, which is the amount of effort or commitment an agent is willing to spend on seeing to its accomplishment. At a finer level the agent’s stance will go beyond an overall position to include a degree of situated autonomy. In this paper, the degree of 1 Assessment of autonomy is either a deliberative process or an automatic association of a stance that might be a historic stance or based on the agent’s personality. In 2000 Proceedings of PRICAI-2000 Workshop on Teams with Adjustable Autonomy, Australia. 2 autonomy will be a measure of the agent’s deliberateness over its autonomy decision. Two useful measures of autonomy beyond the scope of this paper are (1) degree of relative dependence on environmental factors such as other agents, and (2) degree of control (or influence) an agent has over a goal. Generally determining degree of autonomy is more time-consuming than determining an overall position. In our discussion of situated autonomy we will not care whether the goals are internally generated or externally imposed. Action selection generates an action in response to a new situation. An important step in action selection is choosing among possible plans or possible primitive actions. We propose that situated autonomy can be used in this decision. Given that an agent may have several alternative plans and actions to achieve a goal with each alternative appropriate at different commitment level, an agent’s situated autonomy can be used as the context for arbitration. For a cognitive agent, action selection is affected by the frequency of encountering new situations. We describe an agent’s assessment of its situated autonomy that is also affected at varying situation frequencies. At the highest frequency, the agent may produce reflex-like actions. Such agents have no time to account for their situated autonomy. At a relatively fast frequency, the agent may produce reactive actions with minimal time to account for situated autonomy. Such situated autonomy assessment will consider pre-disposition toward the goal. Pre-disposition here is taken as “an evaluative tendency that is expressed by evaluating a particular entity with some degree of favor or disfavor” [3, p. 693]. An agent’s pre-disposition toward a goal is based on semi-permanent beliefs and goals about enabling factors for the goal. Our understanding of pre-disposition is a cognitive function that operates on the agent’s weak beliefs and unmotivated goals. Enabling factors for a goal are subset of the elements of the situation that are either necessary or facilitating conditions for attempting the goal. We consider individual, social, and physical enablers with different origins: (a) entirely endogenous, (b) exogenous and social, and (c) exogenous and physical in nature. The individual enablers are the agent’s competencies. The social enablers are the social influences and motivators. The physical enablers are the physical conditions, physical resources, and physical tools. We will further discuss these types of enablers in a later part of this paper. An agent may have a model of other agents as well as its own. We will use prefixes “Other-” and “Self-” to distinguish between an agent’s model of other agent’s enablers and its own. For instance, Other-Social-Enabler will denote an agent’s model of another agent’s social enablers and Self-Physical-Enabler will denote an agent’s model of its own physical enablers. At slower frequencies, the agent will have more time to assess the quality and quantity of the enabling factors. Situated autonomy at that level will be based on dispositions toward the goal. Perkins, Jay and Tishman [8], define dispositions as “people's tendencies to put their capabilities into action” (p. 75). At yet slower frequencies, the agent will have time to consider going beyond dispositions derived from enabling factors and include motivations. Human motivations are a collection of In 2000 Proceedings of PRICAI-2000 Workshop on Teams with Adjustable Autonomy, Australia. 3 psychogenic needs, which guides behavior [6]. At the slowest frequency, the agent may consider long-term ramifications of its options in order to produce optimal actions. In this paper we consider goal-oriented social agents in domains with relatively fast situation frequency. Such agents may have limited time to consider situated autonomy. They may just have enough time for assessing an overall position. Consider action selection as a linear process where somehow the agent’s action selection has settled on a goal. The next step and the focus of this paper are the agent’s reflections on its autonomy with respect to the goal at hand. Finally, the agent uses the results of its introspection and renders a decision about action(s). Figure 1 summarizes the types of action generated by action selection that is at different frequencies of Situations. Reflex actions are generated without much deliberation for situated autonomy. Other than reflex actions, situated autonomy consideration for actions generated to the right are coarser than for action to the left. Figure 1 Action selection at different frequencies of Situations Given a goal, the agent’s assessment of situated autonomy is a cognitive process that is comprised of several stages, Figure 2. The process begins by the agent’s consideration of pre-dispositions only. If the agent has a habit of disfavoring the goal, it will decide to ignore it for no other reason other than its habit and considers itself Not-autonomous with respect to that goal. The agent who habitually favors the goal and favors itself to be the executor of the goal will consider itself to be Selfautonomous. The agent who habitually favors the goal and favors itself not to be the executor of the goal will lean toward delegation and will consider itself to be Delautonomous. The agent with some more time may consider the goal further and form dispositions toward it. If the agent perceives the goal to be impossible, the agent forms a mental state of Not-autonomous. If the agent perceives that the goal is doable either exclusively by the agent alone or by delegation, it will stop further considerations of situated autonomy. If such an agent solely using its dispositions considers itself to be the executor of the goal, it will consider itself to be Self-autonomous. When we say an agent is autonomous with respect to a goal, we may mean one of two things about its disposition toward the goal. We may mean the agent is self-reliant in the sense that it is not affected by any exogenous sources such as social or physical. Alternatively, we may mean the agent can bring about the desired effect given its access to its Slowest Frequency of Situations f Reflex Pre-disposition Motivated Disposition Disposition Optimized In 2000 Proceedings of PRICAI-2000 Workshop on Teams with Adjustable Autonomy, Australia. 4 exogenous sources such as other agents or resources or it can do it itself. If it considers other agents to be executors of the goal, it will consider itself to be Delautonomous. If the goal is deemed clearly appropriate for delegation due to the agent’s inability to perform the goal itself, the agent is considered Del-autonomous and subsequently a communicative act will be generated. An agent who has formed a disposition toward its goal that has not resulted in determination of either the agent or others being the exclusive executors may further use its motivations. Moreover, motivations can modify a decision that is previously determined based on disposition. We will consider motives to be captured by a policy that produces a preference to favor/disfavor the goal as well as the executor of the goal. If a goal is deemed inappropriate due to the agent’s motivation policy, the initial commitment is revised and the agent is considered to be Not Autonomous with respect to that goal. If a goal is deemed feasible for self-performance due to the agent’s disposition and additionally appropriate due to the agent’s motivation, the agent is considered to Self-autonomous and the goal might be sent to the motoric system. If the agent has not determined exclusive execution, the agent is slated to perform the goal with other agents and its autonomy is classed as Semi-autonomous or Sharedautonomous. Shared-autonomous implies getting assistance from another agent or reliance on some environmental elements such as tools or resources, or offering help to some other agent who will be the primary executioner of the goal. Shared autonomy implies more demand for the agent than semi-autonomy. With semiautonomy, the agent knows it is dependent on outside sources to perform the goal. With shared autonomy the agent knows furthermore that there are one or more agents that complement its autonomy. An example of shared autonomy is shown between a human air traffic controller and a collision-detection agent [5]. Figure 2 Situated Autonomy as part of the process of action selection Situation Goal predisposition disposition