Mixed-initiative assistants are agents that interact seamlessly with humans to extend their probl... more Mixed-initiative assistants are agents that interact seamlessly with humans to extend their problem solving capabilities or provide new capabilities. Developing such agents requires the synergistic integration of many areas of AI, including knowledge representation, problem solving and planning, knowledge acquisition and learning, multi-agent systems, discourse theory, and human-computer interaction. This paper introduces seven aspects of mixed-initiative reasoning (task, control, awareness, communication, personalization, architecture, and evaluation) and discusses them in the context of several state of the art mixed-initiative assistants. The goal is to provide a framework for understanding and comparing existing mixedinitiative assistants, and for the identification of general design principles and methods. Mixed-initiative reasoning concerns the development of collaborative systems where the human and automated agents work together to achieve a common goal in a way that exploits their complementary capabilities. Such systems can either accomplish goals unachievable by the component agents, assuming they work independently, or they can achieve the same goals more effectively. Mixed initiative assumes an efficient, natural interleaving of contributions by users and automated agents that is determined by their relative knowledge and skills and the problemsolving context, rather than by fixed roles, enabling each participant to contribute what it does best, at the appropriate moment. Moreover, dynamic and flexible interaction facilitates adaptation to differences in knowledge, experience, and preferences among different users and to changes in the needs and preferences of individual users over time. Mixed-initiative reasoning represents an important area of Artificial Intelligence because of its potential of achieving both effective human-machine systems where humans interact seamlessly with agents, and multi-agent systems whose capabilities are well above those of the component agents. This area has received considerable attention, as evidenced by a series of workshops (
Various disciplines have examined the many phenomena of metacognition and have produced numerous ... more Various disciplines have examined the many phenomena of metacognition and have produced numerous results, both positive and negative. I discuss some of these aspects of cognition about cognition and the results concerning them from the point of view of the psychologist and the computer scientist, and I attempt to place them in the context of computational theories. I examine metacognition with respect to both problem solving (e.g., planning) and to comprehension (e.g., story understanding) processes of cognition.
Mixed-initiative case replay introduces an active human into the case-based planning process. The... more Mixed-initiative case replay introduces an active human into the case-based planning process. The goals of this novel technique are to utilize the strengths of machine-based case replay to improve human performance and to allow a human to override a machine's abstract representation of an actual domain. The synthesis of these two approaches in a planning domain enables flexible solutions to planning problems. This preliminary research introduces for the first time a human-centered technique into the well-developed planning method of derivational analogy.
Pathological dependency cycles occur in state-space planners when control structures cannot effic... more Pathological dependency cycles occur in state-space planners when control structures cannot efficiently determine a maximal matching for a bipartite operator/binding graph. Without proper search control, the planner will require many computationally expensive backtracks to arrive at a solution. We present a method for improving planning efficiency in the midst of pathological dependency cycles by employing informed resource reallocation in lieu of uninformed backtracking. Empirical studies demonstrate significant improvement in search effort when search control is employed in backtracking. Existing theoretical results suggest that some form of informed resource reallocation can be used to produce an approximately O(n 2.5) solution for many pathological domain classes, as opposed to the O(k n) solution produced in uninformed backtracking.
Complex, dynamic environments present special challenges to autonomous agents. Specifically, agen... more Complex, dynamic environments present special challenges to autonomous agents. Specifically, agents have difficulty when the world does not cooperate with design assumptions. We present an approach to autonomy that seeks to maximize robustness rather than optimality on a specific task. Goaldriven autonomy involves recognizing possibly new problems, explaining what causes the problems and generating goals to solve the problems. We present such a model within the MIDCA cognitive architecture and show that under certain conditions this model outperforms a less flexible approach to handling unexpected events.
A mixed-initiative setting is where both human and machine are intimately involved in the plannin... more A mixed-initiative setting is where both human and machine are intimately involved in the planning process. We have identified a number of challenges that occur for traditional planning frameworks as humans are allowed more latitude. In this paper we will examine the types of unexpected goals that may be given to the underlying planning system and thereby how humans change the way planning must be performed. Users may want to achieve goals in terms of actions as well as states, they may specify goals that vary along a dimension of abstraction and specificity, and they may mix both top-level goals and subgoals when describing what they want a plan to do. We show how the Prodigy planning system has met these challenges when integrated with a force deployment tool called ForMAT and describe what opportunities this poses for a generative planning framework.
Journal of Experimental and Theoretical Artificial Intelligence, Jul 14, 2020
Goal-driven autonomy is an agent model for managing a dynamic environment by reasoning about curr... more Goal-driven autonomy is an agent model for managing a dynamic environment by reasoning about current and potential goals while planning and acting. Since unexpected events and conditions may cause an agent's goals and plans to become invalid or infeasible, an agent with goal-driven autonomy should monitor the environment against its expectations. Designed for dynamic, open, and partially observable environments, such an agent can create new goals or change its existing goals as needed. We present a formalisation of expectations for agents operating in these kinds of environments. Our formalisation includes situations where agents have the capability to sense the environment with some associated costs. We examine agent choices and behaviour in these domains and evaluate multiple approaches for selecting a subset of the agent's sensing actions to execute. The contributions of this work are (1) a specification of different approaches to generating expectations; (2) a formalisation of the autonomy problem that minimises sensing costs; (3) a complexity analysis of the problem; (4) new algorithms for deciding which sensing actions to perform; and (5) empirical results demonstrating the benefit and cost of these approaches.
Humans seek to gain knowledge and structure data by many means including both bottom-up and top-d... more Humans seek to gain knowledge and structure data by many means including both bottom-up and top-down methods. But often, people have a specific purpose to their activity that drives the process, that is, they have particular questions that need answering in support of some broader investigation. These questions often change as answers point in various directions during an investigation, whether the investigation is formal (e.g., scientific, legal, journalistic, or military) or simply an informal browsing of the internet. Here we take a mixed-initiative approach to knowledge discovery, and we present a system called Kyudo that supports the process using a conversational case-based reasoning process. Cases in Kyudo are sequences of knowledge goals or questions that form arcs through a multidimensional knowledge space and that form the core activity in a dialogue between the user and system. As the system gains more experience and therefore more cases, it is able to detect similarity in knowledge goals and prompt the user with additional relevant goals that can short circuit the human reasoning process to minimize tangents or false starts. In this paper we present a distance-based mechanism that reduces the total length of a goal trajectory through guidance that accelerates the human reasoning process and aids effective knowledge discovery.
Metacognition, the ability to monitor and regulate cognition, is important for an agent to adapt ... more Metacognition, the ability to monitor and regulate cognition, is important for an agent to adapt to novel situations and fix discrepancies in its knowledge base. In this paper, we discuss two different approaches to implementing metacognition in artificial systems: internally and externally. In the internal approach, metacognition is built into the agent, and thus, is combined with its cognitive reasoning abilities. The same Knowledge Base (KB) is fully shared between the metacognitive and cognitive processes of the artificial system. In the external metacognition approach, only portions of the agent's KB are shared between the agent and the external metacognition. We describe the implementation of external metacognition using our own Metacognitive Loop (MCL2) and internal metacognition using active logic in the context of a dialog agent called Alfred. We discuss how the two systems handle long pauses in dialog and compare the pros and cons of each. Our experiments show that for a system with time-related expectations, it is more efficient to use interleaved metacognition rather than an external metacognition module as the internal metacognition has access to the entire KB of the agent.
ABSTRACT Abstract Inmultiagent distributed planning systems, agents can act cooperatively toward ... more ABSTRACT Abstract Inmultiagent distributed planning systems, agents can act cooperatively toward ,a common ,goal. Using ,multiagent goal transformations, the planning process can be distributed over an appropriate number ,of autonomous agents. Each ,agent will attempt to plan ,its part of the problem,using its share of the ,knowledge. We will ,show how,the distribution of knowledge affects the planning performance,of the ,system. The factors we will ,examine inthis process include the number ,of agents ,used in the system, which agent will initiate the planning process, and what knowledge, if any, will be duplicated among the agents. We ,will use ,examples ,from ,the ,logistics transportation domain ,to demonstrate ,how ,these factors affect performance.
For a case-based application with a sizable case library, a common practice is to retrieve a set ... more For a case-based application with a sizable case library, a common practice is to retrieve a set of cases and then reduce the set with a domain-specific similarity metric. This research investigates an alternative technique that constructs a number of two-dimensional (or multi-dimensional) indices using the principle of elaboration and specialization. By storing cases with these indices, we reduce the size of the retrieved candidate set and, in many instances, fetch a single case. This paper describes a case-based reasoner called SMIRKS. We investigate the retrieval performance by comparing linear search, 1-dimensional indexing, and 2dimensional indexing. The improvement in performance with dimensional indexing is found to be significant, especially in terms of the size of the retrieved candidate set. This paper describes the implementation of SMIRKS, presents the results of evaluation, and discusses some ideas on future applications that can utilize this technique.
... This particular heuristic, however, does not generalize across all planning domains. Fig. ...... more ... This particular heuristic, however, does not generalize across all planning domains. Fig. ... Such representational change facilitates the complexity analysis of the argument substitution scheme. Practical Adaptation with Argument Substitution ...
Workflow management systems (WfMS) allow multiple agents to work towards achieving a common goal ... more Workflow management systems (WfMS) allow multiple agents to work towards achieving a common goal by facilitating communication between them. This paper discusses the distinctive characteristics of portal-based WfMS and considers the utility of using techniques employed in other WfMS environments in this domain. Specifically, the idea of constructing workflows by applying artificial intelligence planning techniques to a userspecified goal is explored.
International Conference on Automated Planning and Scheduling, Jun 5, 2005
Mixed-initiative planning systems attempt to integrate human and AI planners so that the synthesi... more Mixed-initiative planning systems attempt to integrate human and AI planners so that the synthesis results in high quality plans. In the AI community, the dominant model of planning is search. In state-space planning, search consists of backward and forward chaining through the effects and preconditions of operator representations. Although search is an acceptable mechanism to use in performing automated planning, we present an alternative model to present to the user at the interface of a mixed-initiative planning system. That is we propose to model planning as a goal manipulation task. Here planning involves moving goals through a hyperspace in order to reach equilibrium between available resources and the constraints of a dynamic environment. The users can establish and "steer" goals through a visual representation of the planning domain. They can associate resources with particular goals and shift goals along various dimensions in response to changing conditions as well as change the structure of previous plans. Users need not know details of the underlying technology, even when search is used within. Here we empirically examine user performance under both alternatives and see that many users do better with the alternative model.
Mixed-initiative assistants are agents that interact seamlessly with humans to extend their probl... more Mixed-initiative assistants are agents that interact seamlessly with humans to extend their problem solving capabilities or provide new capabilities. Developing such agents requires the synergistic integration of many areas of AI, including knowledge representation, problem solving and planning, knowledge acquisition and learning, multi-agent systems, discourse theory, and human-computer interaction. This paper introduces seven aspects of mixed-initiative reasoning (task, control, awareness, communication, personalization, architecture, and evaluation) and discusses them in the context of several state of the art mixed-initiative assistants. The goal is to provide a framework for understanding and comparing existing mixedinitiative assistants, and for the identification of general design principles and methods. Mixed-initiative reasoning concerns the development of collaborative systems where the human and automated agents work together to achieve a common goal in a way that exploits their complementary capabilities. Such systems can either accomplish goals unachievable by the component agents, assuming they work independently, or they can achieve the same goals more effectively. Mixed initiative assumes an efficient, natural interleaving of contributions by users and automated agents that is determined by their relative knowledge and skills and the problemsolving context, rather than by fixed roles, enabling each participant to contribute what it does best, at the appropriate moment. Moreover, dynamic and flexible interaction facilitates adaptation to differences in knowledge, experience, and preferences among different users and to changes in the needs and preferences of individual users over time. Mixed-initiative reasoning represents an important area of Artificial Intelligence because of its potential of achieving both effective human-machine systems where humans interact seamlessly with agents, and multi-agent systems whose capabilities are well above those of the component agents. This area has received considerable attention, as evidenced by a series of workshops (
Various disciplines have examined the many phenomena of metacognition and have produced numerous ... more Various disciplines have examined the many phenomena of metacognition and have produced numerous results, both positive and negative. I discuss some of these aspects of cognition about cognition and the results concerning them from the point of view of the psychologist and the computer scientist, and I attempt to place them in the context of computational theories. I examine metacognition with respect to both problem solving (e.g., planning) and to comprehension (e.g., story understanding) processes of cognition.
Mixed-initiative case replay introduces an active human into the case-based planning process. The... more Mixed-initiative case replay introduces an active human into the case-based planning process. The goals of this novel technique are to utilize the strengths of machine-based case replay to improve human performance and to allow a human to override a machine's abstract representation of an actual domain. The synthesis of these two approaches in a planning domain enables flexible solutions to planning problems. This preliminary research introduces for the first time a human-centered technique into the well-developed planning method of derivational analogy.
Pathological dependency cycles occur in state-space planners when control structures cannot effic... more Pathological dependency cycles occur in state-space planners when control structures cannot efficiently determine a maximal matching for a bipartite operator/binding graph. Without proper search control, the planner will require many computationally expensive backtracks to arrive at a solution. We present a method for improving planning efficiency in the midst of pathological dependency cycles by employing informed resource reallocation in lieu of uninformed backtracking. Empirical studies demonstrate significant improvement in search effort when search control is employed in backtracking. Existing theoretical results suggest that some form of informed resource reallocation can be used to produce an approximately O(n 2.5) solution for many pathological domain classes, as opposed to the O(k n) solution produced in uninformed backtracking.
Complex, dynamic environments present special challenges to autonomous agents. Specifically, agen... more Complex, dynamic environments present special challenges to autonomous agents. Specifically, agents have difficulty when the world does not cooperate with design assumptions. We present an approach to autonomy that seeks to maximize robustness rather than optimality on a specific task. Goaldriven autonomy involves recognizing possibly new problems, explaining what causes the problems and generating goals to solve the problems. We present such a model within the MIDCA cognitive architecture and show that under certain conditions this model outperforms a less flexible approach to handling unexpected events.
A mixed-initiative setting is where both human and machine are intimately involved in the plannin... more A mixed-initiative setting is where both human and machine are intimately involved in the planning process. We have identified a number of challenges that occur for traditional planning frameworks as humans are allowed more latitude. In this paper we will examine the types of unexpected goals that may be given to the underlying planning system and thereby how humans change the way planning must be performed. Users may want to achieve goals in terms of actions as well as states, they may specify goals that vary along a dimension of abstraction and specificity, and they may mix both top-level goals and subgoals when describing what they want a plan to do. We show how the Prodigy planning system has met these challenges when integrated with a force deployment tool called ForMAT and describe what opportunities this poses for a generative planning framework.
Journal of Experimental and Theoretical Artificial Intelligence, Jul 14, 2020
Goal-driven autonomy is an agent model for managing a dynamic environment by reasoning about curr... more Goal-driven autonomy is an agent model for managing a dynamic environment by reasoning about current and potential goals while planning and acting. Since unexpected events and conditions may cause an agent's goals and plans to become invalid or infeasible, an agent with goal-driven autonomy should monitor the environment against its expectations. Designed for dynamic, open, and partially observable environments, such an agent can create new goals or change its existing goals as needed. We present a formalisation of expectations for agents operating in these kinds of environments. Our formalisation includes situations where agents have the capability to sense the environment with some associated costs. We examine agent choices and behaviour in these domains and evaluate multiple approaches for selecting a subset of the agent's sensing actions to execute. The contributions of this work are (1) a specification of different approaches to generating expectations; (2) a formalisation of the autonomy problem that minimises sensing costs; (3) a complexity analysis of the problem; (4) new algorithms for deciding which sensing actions to perform; and (5) empirical results demonstrating the benefit and cost of these approaches.
Humans seek to gain knowledge and structure data by many means including both bottom-up and top-d... more Humans seek to gain knowledge and structure data by many means including both bottom-up and top-down methods. But often, people have a specific purpose to their activity that drives the process, that is, they have particular questions that need answering in support of some broader investigation. These questions often change as answers point in various directions during an investigation, whether the investigation is formal (e.g., scientific, legal, journalistic, or military) or simply an informal browsing of the internet. Here we take a mixed-initiative approach to knowledge discovery, and we present a system called Kyudo that supports the process using a conversational case-based reasoning process. Cases in Kyudo are sequences of knowledge goals or questions that form arcs through a multidimensional knowledge space and that form the core activity in a dialogue between the user and system. As the system gains more experience and therefore more cases, it is able to detect similarity in knowledge goals and prompt the user with additional relevant goals that can short circuit the human reasoning process to minimize tangents or false starts. In this paper we present a distance-based mechanism that reduces the total length of a goal trajectory through guidance that accelerates the human reasoning process and aids effective knowledge discovery.
Metacognition, the ability to monitor and regulate cognition, is important for an agent to adapt ... more Metacognition, the ability to monitor and regulate cognition, is important for an agent to adapt to novel situations and fix discrepancies in its knowledge base. In this paper, we discuss two different approaches to implementing metacognition in artificial systems: internally and externally. In the internal approach, metacognition is built into the agent, and thus, is combined with its cognitive reasoning abilities. The same Knowledge Base (KB) is fully shared between the metacognitive and cognitive processes of the artificial system. In the external metacognition approach, only portions of the agent's KB are shared between the agent and the external metacognition. We describe the implementation of external metacognition using our own Metacognitive Loop (MCL2) and internal metacognition using active logic in the context of a dialog agent called Alfred. We discuss how the two systems handle long pauses in dialog and compare the pros and cons of each. Our experiments show that for a system with time-related expectations, it is more efficient to use interleaved metacognition rather than an external metacognition module as the internal metacognition has access to the entire KB of the agent.
ABSTRACT Abstract Inmultiagent distributed planning systems, agents can act cooperatively toward ... more ABSTRACT Abstract Inmultiagent distributed planning systems, agents can act cooperatively toward ,a common ,goal. Using ,multiagent goal transformations, the planning process can be distributed over an appropriate number ,of autonomous agents. Each ,agent will attempt to plan ,its part of the problem,using its share of the ,knowledge. We will ,show how,the distribution of knowledge affects the planning performance,of the ,system. The factors we will ,examine inthis process include the number ,of agents ,used in the system, which agent will initiate the planning process, and what knowledge, if any, will be duplicated among the agents. We ,will use ,examples ,from ,the ,logistics transportation domain ,to demonstrate ,how ,these factors affect performance.
For a case-based application with a sizable case library, a common practice is to retrieve a set ... more For a case-based application with a sizable case library, a common practice is to retrieve a set of cases and then reduce the set with a domain-specific similarity metric. This research investigates an alternative technique that constructs a number of two-dimensional (or multi-dimensional) indices using the principle of elaboration and specialization. By storing cases with these indices, we reduce the size of the retrieved candidate set and, in many instances, fetch a single case. This paper describes a case-based reasoner called SMIRKS. We investigate the retrieval performance by comparing linear search, 1-dimensional indexing, and 2dimensional indexing. The improvement in performance with dimensional indexing is found to be significant, especially in terms of the size of the retrieved candidate set. This paper describes the implementation of SMIRKS, presents the results of evaluation, and discusses some ideas on future applications that can utilize this technique.
... This particular heuristic, however, does not generalize across all planning domains. Fig. ...... more ... This particular heuristic, however, does not generalize across all planning domains. Fig. ... Such representational change facilitates the complexity analysis of the argument substitution scheme. Practical Adaptation with Argument Substitution ...
Workflow management systems (WfMS) allow multiple agents to work towards achieving a common goal ... more Workflow management systems (WfMS) allow multiple agents to work towards achieving a common goal by facilitating communication between them. This paper discusses the distinctive characteristics of portal-based WfMS and considers the utility of using techniques employed in other WfMS environments in this domain. Specifically, the idea of constructing workflows by applying artificial intelligence planning techniques to a userspecified goal is explored.
International Conference on Automated Planning and Scheduling, Jun 5, 2005
Mixed-initiative planning systems attempt to integrate human and AI planners so that the synthesi... more Mixed-initiative planning systems attempt to integrate human and AI planners so that the synthesis results in high quality plans. In the AI community, the dominant model of planning is search. In state-space planning, search consists of backward and forward chaining through the effects and preconditions of operator representations. Although search is an acceptable mechanism to use in performing automated planning, we present an alternative model to present to the user at the interface of a mixed-initiative planning system. That is we propose to model planning as a goal manipulation task. Here planning involves moving goals through a hyperspace in order to reach equilibrium between available resources and the constraints of a dynamic environment. The users can establish and "steer" goals through a visual representation of the planning domain. They can associate resources with particular goals and shift goals along various dimensions in response to changing conditions as well as change the structure of previous plans. Users need not know details of the underlying technology, even when search is used within. Here we empirically examine user performance under both alternatives and see that many users do better with the alternative model.
Uploads
Papers by Michael Cox