Citation Context

...er we have proposed an architecture that includes (re) planning in BDI agents. The proposed architecture describes how to integrate a real-time planner with replanning capability in the current BDI architecture. Replanning capability is important for reactive behaviour. Key words: BDI agent, AI planning 1. INTRODUCTION Intelligent software agents are powerful tools in today’s modern software systems. They have been deployed in complex and dynamic hostile environment and even in unknown environments. Research in intelligent agents is very active and progressive. According to Russell and Norvig [14] 2 Debdeep Banerjee, Jeffrey Tweedale “An agent can be anything that perceives its environment through sensors prior to acting upon the environment through actuators”. Wooldridge defines an agent as “a computer system capable of autonomous action in a given environment in order to meet its design objectives” [1]. Jennings adds four main properties to these definitions which are: autonomy, reactivity, proactivity and social ability [2]. Another important aspect to consider is its environment that an agent operates [14]. The agent’s characteristics and capabilities also depend on the environment...

Citation Context

...and dynamic hostile environment and even in unknown environments. Research in intelligent agents is very active and progressive. According to Russell and Norvig [14] 2 Debdeep Banerjee, Jeffrey Tweedale “An agent can be anything that perceives its environment through sensors prior to acting upon the environment through actuators”. Wooldridge defines an agent as “a computer system capable of autonomous action in a given environment in order to meet its design objectives” [1]. Jennings adds four main properties to these definitions which are: autonomy, reactivity, proactivity and social ability [2]. Another important aspect to consider is its environment that an agent operates [14]. The agent’s characteristics and capabilities also depend on the environment, because agent has to interact with its environment. Agent’s environment has been divided between Static and Dynamic environment, fully Observable and partially Observable environment, Continuous and Discrete environment, Deterministic and Nondeterministic environment. There are mainly three common categories of architectures used to design intelligent agents: the Brooks subsumption architecture [17], Bratmans’ Belief-Desires-Intensi...

Citation Context

...t will help it achieve its goals. Practical reasoning is used to decide what to do (deliberation) and how to do it (means-end-analysis) [1]. Planning is intrinsic for any intelligent behaviour. Humans tend to plan most events they contribute to in the real world. The planning process may not always be visible, especially when those actions require routine skills, rule based tasks and procedures performed by subject matter experts. Humans’ (re) use the rules, skills and knowledge stored in memory (plans that need to be embodied into agents) to achieve tasks they may have previously encountered [16]. We tend to plan for situations that are new, complex or critical. Planning is a costly process in the terms of time and computation. The motivation for Intelligent Agents is to personify human capabilities, so they can be used in place of humans and how they achieve the given goal by acting rationally in their environment. The main part of a rational act is that of practical reasoning. So the planning is the part of the practical reasoning as it describes a set of actions to achieve a goal [4]. This paper is organized in four parts. Section two describes the problems relating to planning in ...

Citation Context

...UCTION Intelligent software agents are powerful tools in today’s modern software systems. They have been deployed in complex and dynamic hostile environment and even in unknown environments. Research in intelligent agents is very active and progressive. According to Russell and Norvig [14] 2 Debdeep Banerjee, Jeffrey Tweedale “An agent can be anything that perceives its environment through sensors prior to acting upon the environment through actuators”. Wooldridge defines an agent as “a computer system capable of autonomous action in a given environment in order to meet its design objectives” [1]. Jennings adds four main properties to these definitions which are: autonomy, reactivity, proactivity and social ability [2]. Another important aspect to consider is its environment that an agent operates [14]. The agent’s characteristics and capabilities also depend on the environment, because agent has to interact with its environment. Agent’s environment has been divided between Static and Dynamic environment, fully Observable and partially Observable environment, Continuous and Discrete environment, Deterministic and Nondeterministic environment. There are mainly three common categories o...

Citation Context

...pecific task (such as monitoring the communication network and fault diagnosis). Artificial Intelligent agents only need to decide how to achieve its goals by using knowledge about the environment and knowledge about its capabilities and measures. Within this process the agent should identify any sub-tasks to be achieved given the goal and suitable plans to achieve them. To succeed, autonomous agents must have a planning component capable of synthesizing its own course of actions from within the environment it resides. There are agent architectures (such as RETSINA [19], PROPICE [20], CYPRESS [21], INTERRAP [22], TAIPE [23] etc.) that incorporate a planning component as part of the agent architecture. These systems implement different architecture to incorporate the planning module. Our proposed architecture extends the BDI agent architecture with online planning capabilities which will be handled within main BDI loop rather than accessed as an external component. 3. A FLEXIBLE PLANNING ARCHITECTURE FOR BDI AGENTS To provide reactive behaviour to the BDI agents a new architecture has been proposed (Fig 2). This architecture is an extension of the BDI 4 Debdeep Banerjee, Jeffrey Tweedal...

Citation Context

...ty, proactivity and social ability [2]. Another important aspect to consider is its environment that an agent operates [14]. The agent’s characteristics and capabilities also depend on the environment, because agent has to interact with its environment. Agent’s environment has been divided between Static and Dynamic environment, fully Observable and partially Observable environment, Continuous and Discrete environment, Deterministic and Nondeterministic environment. There are mainly three common categories of architectures used to design intelligent agents: the Brooks subsumption architecture [17], Bratmans’ Belief-Desires-Intension (BDI) architecture [4] and the layered model [18]. The BDI model was derived from the model of human practical reasoning system [4] based on rational agents that conduct actions that will help it achieve its goals. Practical reasoning is used to decide what to do (deliberation) and how to do it (means-end-analysis) [1]. Planning is intrinsic for any intelligent behaviour. Humans tend to plan most events they contribute to in the real world. The planning process may not always be visible, especially when those actions require routine skills, rule based tasks...

Citation Context

...ons [10, 11]. In the dynamic environment the environment changes very fast. An agent situated in this dynamic environment needs to react to these changes. For this highly reactive behaviour agent may not need to synthesis a full plan for achieving goal. This reactivity can be achieved by incorporating anytime algorithm based planner [7]. In anytime based planner a planner can be interrupted at anytime and planner always have some executable plan as the result [7, 8]. The main problem in this kind of planner is to guarantee the quality of the resultant plan. On the other hand genetic algorithm [12, 13] can also be implemented so at any point of time agent can have an executable plan. The Sub-goal Deliberation module can be compared to the plan library of the BDI agent architecture. It can contain the domain specific knowledge in the form of predefined task decomposition. Only difference would be instead of producing an executable plan it will produce abstract level tasks as sub goals. We can incorporate different strategies, such as decision theoretic approach, case-based approach, knowledge-based approach, for Sub-goal Deliberation. Sub-goal Deliberation process can be modeled as planning ...

Citation Context

.... The proposed architecture introduced online planning with replanning capability in BDI agent architecture. This architecture can use the domain knowledge for Sub-goal Deliberation and provide flexibility for different types of dynamic domains. This architecture can also be extent in the cases where the environment is not fully observable and changes frequently in random manner. The implementation phase has four main steps. The first step is to find a common representation of the planning problem between the planner and the agent architecture. We will use JACK as our BDI implementation. JACK [15] is a BDI based commercial strength multi-agent based software development framework based on JAVA. It is developed by the Agent Reactive (Re) Planning Agents in a Dynamic Environment 9 Oriented Software. JACK provides a high performance, lightweight implementation of BDI architecture. It is an agent oriented programming extension of JAVA. The second step would be interfacing a external planner with the JACK agent or incorporating an planning algorithm within JACK. The next step would be implementing a State Change Monitor in context of JACK system. Last step would be to extend the planner wit...

Citation Context

... agent is designed to do a specific task (such as monitoring the communication network and fault diagnosis). Artificial Intelligent agents only need to decide how to achieve its goals by using knowledge about the environment and knowledge about its capabilities and measures. Within this process the agent should identify any sub-tasks to be achieved given the goal and suitable plans to achieve them. To succeed, autonomous agents must have a planning component capable of synthesizing its own course of actions from within the environment it resides. There are agent architectures (such as RETSINA [19], PROPICE [20], CYPRESS [21], INTERRAP [22], TAIPE [23] etc.) that incorporate a planning component as part of the agent architecture. These systems implement different architecture to incorporate the planning module. Our proposed architecture extends the BDI agent architecture with online planning capabilities which will be handled within main BDI loop rather than accessed as an external component. 3. A FLEXIBLE PLANNING ARCHITECTURE FOR BDI AGENTS To provide reactive behaviour to the BDI agents a new architecture has been proposed (Fig 2). This architecture is an extension of the BDI 4 Debde...

Citation Context

...vironment that an agent operates [14]. The agent’s characteristics and capabilities also depend on the environment, because agent has to interact with its environment. Agent’s environment has been divided between Static and Dynamic environment, fully Observable and partially Observable environment, Continuous and Discrete environment, Deterministic and Nondeterministic environment. There are mainly three common categories of architectures used to design intelligent agents: the Brooks subsumption architecture [17], Bratmans’ Belief-Desires-Intension (BDI) architecture [4] and the layered model [18]. The BDI model was derived from the model of human practical reasoning system [4] based on rational agents that conduct actions that will help it achieve its goals. Practical reasoning is used to decide what to do (deliberation) and how to do it (means-end-analysis) [1]. Planning is intrinsic for any intelligent behaviour. Humans tend to plan most events they contribute to in the real world. The planning process may not always be visible, especially when those actions require routine skills, rule based tasks and procedures performed by subject matter experts. Humans’ (re) use the rules, skill...

Citation Context

...uch as monitoring the communication network and fault diagnosis). Artificial Intelligent agents only need to decide how to achieve its goals by using knowledge about the environment and knowledge about its capabilities and measures. Within this process the agent should identify any sub-tasks to be achieved given the goal and suitable plans to achieve them. To succeed, autonomous agents must have a planning component capable of synthesizing its own course of actions from within the environment it resides. There are agent architectures (such as RETSINA [19], PROPICE [20], CYPRESS [21], INTERRAP [22], TAIPE [23] etc.) that incorporate a planning component as part of the agent architecture. These systems implement different architecture to incorporate the planning module. Our proposed architecture extends the BDI agent architecture with online planning capabilities which will be handled within main BDI loop rather than accessed as an external component. 3. A FLEXIBLE PLANNING ARCHITECTURE FOR BDI AGENTS To provide reactive behaviour to the BDI agents a new architecture has been proposed (Fig 2). This architecture is an extension of the BDI 4 Debdeep Banerjee, Jeffrey Tweedale architecture ...

Citation Context

...ks if the plan is still consistent with the current world after execution of each action. To check consistency it will check if all the preconditions of the actions that are still to be executed at the next step are true and the goal is still achievable but not achieved yet [14]. For the planning module we assume the environment is fully observable. To incorporate the planning module we need to find a common representation of the actions between the chosen planner and the agent architecture. There are similarities between the BDI architecture and HTN (Hierarchical Task Network) based planning [5]. A wrapper can be created that maps the BDI agent syntax to the HTN based planner syntax similar to [6]. For replanning capabilities we can implement a replanning module on top of the planner. There are different options exist for replanning. First option would be introducing replanning algorithm [8, 9] which can start replanning by backtracking from the point where the plan fails and choose 8 Debdeep Banerjee, Jeffrey Tweedale alternative path. Replanning can also be done by plan refinement technique where in the case of a failed plan refinement technique replaces the failed actions with the...

Citation Context

...le we assume the environment is fully observable. To incorporate the planning module we need to find a common representation of the actions between the chosen planner and the agent architecture. There are similarities between the BDI architecture and HTN (Hierarchical Task Network) based planning [5]. A wrapper can be created that maps the BDI agent syntax to the HTN based planner syntax similar to [6]. For replanning capabilities we can implement a replanning module on top of the planner. There are different options exist for replanning. First option would be introducing replanning algorithm [8, 9] which can start replanning by backtracking from the point where the plan fails and choose 8 Debdeep Banerjee, Jeffrey Tweedale alternative path. Replanning can also be done by plan refinement technique where in the case of a failed plan refinement technique replaces the failed actions with the alternate actions [10, 11]. In the dynamic environment the environment changes very fast. An agent situated in this dynamic environment needs to react to these changes. For this highly reactive behaviour agent may not need to synthesis a full plan for achieving goal. This reactivity can be achieved by i...

Citation Context

...oring the communication network and fault diagnosis). Artificial Intelligent agents only need to decide how to achieve its goals by using knowledge about the environment and knowledge about its capabilities and measures. Within this process the agent should identify any sub-tasks to be achieved given the goal and suitable plans to achieve them. To succeed, autonomous agents must have a planning component capable of synthesizing its own course of actions from within the environment it resides. There are agent architectures (such as RETSINA [19], PROPICE [20], CYPRESS [21], INTERRAP [22], TAIPE [23] etc.) that incorporate a planning component as part of the agent architecture. These systems implement different architecture to incorporate the planning module. Our proposed architecture extends the BDI agent architecture with online planning capabilities which will be handled within main BDI loop rather than accessed as an external component. 3. A FLEXIBLE PLANNING ARCHITECTURE FOR BDI AGENTS To provide reactive behaviour to the BDI agents a new architecture has been proposed (Fig 2). This architecture is an extension of the BDI 4 Debdeep Banerjee, Jeffrey Tweedale architecture (Fig 1) as W...

Citation Context

...of a rational act is that of practical reasoning. So the planning is the part of the practical reasoning as it describes a set of actions to achieve a goal [4]. This paper is organized in four parts. Section two describes the problems relating to planning in BDI architecture and the motivation of the proposed architecture. Next section describes the proposed architecture and section four contains the research methodology. Section five concludes the paper with the future directions. 2. MOTIVATION BDI agents use a plan library or predefined set of plans, instead of planning from first principle [3]. When an agent commits to an intension, it Reactive (Re) Planning Agents in a Dynamic Environment 3 looks through its plan library for feasible plan, which is executed in order to achieve the goal. If the plan fails and a suitable alternative can’t be identified then the agent fails to achieve the goal [3]. The main bottleneck of most BDI architectures is the plan library. Library agents are predominantly deterministic in nature, because all of its behaviour is hard coded into the library. The knowledge about how to achieve a specific task must be explicitly captured as plans in the library b...

Citation Context

...ons [10, 11]. In the dynamic environment the environment changes very fast. An agent situated in this dynamic environment needs to react to these changes. For this highly reactive behaviour agent may not need to synthesis a full plan for achieving goal. This reactivity can be achieved by incorporating anytime algorithm based planner [7]. In anytime based planner a planner can be interrupted at anytime and planner always have some executable plan as the result [7, 8]. The main problem in this kind of planner is to guarantee the quality of the resultant plan. On the other hand genetic algorithm [12, 13] can also be implemented so at any point of time agent can have an executable plan. The Sub-goal Deliberation module can be compared to the plan library of the BDI agent architecture. It can contain the domain specific knowledge in the form of predefined task decomposition. Only difference would be instead of producing an executable plan it will produce abstract level tasks as sub goals. We can incorporate different strategies, such as decision theoretic approach, case-based approach, knowledge-based approach, for Sub-goal Deliberation. Sub-goal Deliberation process can be modeled as planning ...

Citation Context

...created that maps the BDI agent syntax to the HTN based planner syntax similar to [6]. For replanning capabilities we can implement a replanning module on top of the planner. There are different options exist for replanning. First option would be introducing replanning algorithm [8, 9] which can start replanning by backtracking from the point where the plan fails and choose 8 Debdeep Banerjee, Jeffrey Tweedale alternative path. Replanning can also be done by plan refinement technique where in the case of a failed plan refinement technique replaces the failed actions with the alternate actions [10, 11]. In the dynamic environment the environment changes very fast. An agent situated in this dynamic environment needs to react to these changes. For this highly reactive behaviour agent may not need to synthesis a full plan for achieving goal. This reactivity can be achieved by incorporating anytime algorithm based planner [7]. In anytime based planner a planner can be interrupted at anytime and planner always have some executable plan as the result [7, 8]. The main problem in this kind of planner is to guarantee the quality of the resultant plan. On the other hand genetic algorithm [12, 13] can...

Citation Context

...pect to consider is its environment that an agent operates [14]. The agent’s characteristics and capabilities also depend on the environment, because agent has to interact with its environment. Agent’s environment has been divided between Static and Dynamic environment, fully Observable and partially Observable environment, Continuous and Discrete environment, Deterministic and Nondeterministic environment. There are mainly three common categories of architectures used to design intelligent agents: the Brooks subsumption architecture [17], Bratmans’ Belief-Desires-Intension (BDI) architecture [4] and the layered model [18]. The BDI model was derived from the model of human practical reasoning system [4] based on rational agents that conduct actions that will help it achieve its goals. Practical reasoning is used to decide what to do (deliberation) and how to do it (means-end-analysis) [1]. Planning is intrinsic for any intelligent behaviour. Humans tend to plan most events they contribute to in the real world. The planning process may not always be visible, especially when those actions require routine skills, rule based tasks and procedures performed by subject matter experts. Humans...

Citation Context

...stency it will check if all the preconditions of the actions that are still to be executed at the next step are true and the goal is still achievable but not achieved yet [14]. For the planning module we assume the environment is fully observable. To incorporate the planning module we need to find a common representation of the actions between the chosen planner and the agent architecture. There are similarities between the BDI architecture and HTN (Hierarchical Task Network) based planning [5]. A wrapper can be created that maps the BDI agent syntax to the HTN based planner syntax similar to [6]. For replanning capabilities we can implement a replanning module on top of the planner. There are different options exist for replanning. First option would be introducing replanning algorithm [8, 9] which can start replanning by backtracking from the point where the plan fails and choose 8 Debdeep Banerjee, Jeffrey Tweedale alternative path. Replanning can also be done by plan refinement technique where in the case of a failed plan refinement technique replaces the failed actions with the alternate actions [10, 11]. In the dynamic environment the environment changes very fast. An agent situ...

Citation Context

...le we assume the environment is fully observable. To incorporate the planning module we need to find a common representation of the actions between the chosen planner and the agent architecture. There are similarities between the BDI architecture and HTN (Hierarchical Task Network) based planning [5]. A wrapper can be created that maps the BDI agent syntax to the HTN based planner syntax similar to [6]. For replanning capabilities we can implement a replanning module on top of the planner. There are different options exist for replanning. First option would be introducing replanning algorithm [8, 9] which can start replanning by backtracking from the point where the plan fails and choose 8 Debdeep Banerjee, Jeffrey Tweedale alternative path. Replanning can also be done by plan refinement technique where in the case of a failed plan refinement technique replaces the failed actions with the alternate actions [10, 11]. In the dynamic environment the environment changes very fast. An agent situated in this dynamic environment needs to react to these changes. For this highly reactive behaviour agent may not need to synthesis a full plan for achieving goal. This reactivity can be achieved by i...