Automated problem resolution learning for intent manager

Abstract

A method performed by an intent manager that controls an environment based on an intent for the environment includes receiving information about the environment. Based on the information, the method identifies a problem in the environment that needs to be solved to implement the intent, and obtains an efficiency map associated with the problem. The efficiency map indicates an efficiency score of the suggested agent when generating a suggested action for the environment to solve the problem. Based on the efficiency score, a suggestion agent is invoked to generate a suggested action for the environment to solve the problem. The method determines whether a suggested action generated by a suggested agent is selected for implementation on the environment; and adjusting an efficiency score of the suggested agent based on whether the suggested action is selected for implementation on the environment.

Description

Technical Field

[0001] The present disclosure relates to systems and methods for controlling an environment, and more particularly, to intent management functionality, and more particularly, to systems and methods for controlling an environment using zero-touch control. Background Art

[0002] Intent Managers (IMs) provide zero-touch control of the environment. "Zero-touch control" refers to systems or processes that can be automatically configured and managed without human intervention. The term is often used in the context of networking and information technology, where zero-touch provisioning allows new devices to be added to the network without manual configuration. For example, see "Intent-Based Networking—Concepts and Definitions" published by the Internet Engineering Task Force (IETF). This can be accomplished using preconfigured settings, remote management tools, or other automated technologies. The goal of zero-touch control is to reduce the time and resources required to manage and maintain complex systems.

[0003] refer to Figure 1 Intent manager 10 is controlled by one or more intents provided by operator 12. Intent manager 10 controls one or more environments 20. Controlling environment 20 is accomplished by observing environment 20, reasoning based on a combination of perceptions and prior knowledge of environment 20, and taking actions on environment 20. These steps can be performed in a closed loop, where environment 20 changes based on the actions taken. The overall purpose of the intent manager is to satisfy the intent of the operator.

[0004] In this context, the term "intent" refers to the formal specification of all expectations, including requirements, goals, and constraints imposed on an environment. When the environment is a communication system, expectations can refer to operational metrics or parameters. For example, expectations can be expressed as minimum requirements, such as "at least 95% of URLLC (Ultra-Reliable Low Latency Communication) users should experience a maximum latency of 20 milliseconds," or "at least 80% of users of a conversational video service should have a minimum QoE (Quality of Experience) of 4.0," or "the system's energy consumption should be kept to a minimum."

[0005] exist Figure 1 The intent manager 10 shown at a high level in FIG can be used as shown in FIG. Figure 2 As shown, the intent manager 10 generally includes a knowledge base 32 and a plurality of agents, including one or more data grounding agents 34a, a suggestion agent 34b, an evaluation agent 34c, and an actuator agent 34d.

[0006] One or more intents are sent to the intent manager 10. Each expectation in the intent can be viewed as a key performance indicator (KPI) that needs to be met, called a target KPI. Raw data is exposed from the environment and processed by the data anchor agent 34a. The data anchor agent 34a converts the raw data into measured KPIs. The target KPI and the measured KPI can be compared, and the difference becomes the problem or goal that the intent manager needs to meet. For example, the target KPI is "maximum 20 millisecond latency", but the measured KPI is "30 millisecond latency". One or more suggestion agents 34b are responsible for suggesting actions that will solve the problem. The evaluation agent 34c evaluates whether the suggested action is good or not. Finally, the actuator agent 34d executes the action on the environment 20 under control.

[0007] When the general Figure 2 When the structure is applied to a communication network such as a mobile communication network or a wireless communication network, the controlled environment 20 corresponds to the communication network, and the operator can be a human network operator. For scalability reasons, the communication network can be divided into multiple domains (i.e., environments). In addition, multiple intent managers can be arranged in a hierarchical structure. Generally speaking, the controlled environment can be part of the communication network or another intent manager. The operator can be a human network operator or another intent manager. Summary of the Invention

[0008] A method performed by an intent manager for controlling an environment based on an intent for the environment includes receiving information about the environment. Based on the information about the environment, the method identifies a problem in the environment that needs to be solved to achieve the intent; and obtains an efficiency graph associated with the problem. The efficiency graph indicates an efficiency score of a first suggestion agent when generating a suggested action for the environment to solve the problem. Based on the efficiency score, the method calls the first suggestion agent to generate a suggested action for the environment to solve the problem. The method determines whether the suggested action generated by the first suggestion agent is selected for implementation on the environment; and based on whether the suggested action generated by the first suggestion agent is selected for implementation on the environment, adjusts the efficiency score of the first suggestion agent.

[0009] Adjusting the efficiency score of the first suggestion agent may include increasing the efficiency score of the first suggestion agent in response to the suggested action generated by the first suggestion agent being selected for implementation on the environment, and decreasing the efficiency score of the first suggestion agent in response to the suggested action generated by the first suggestion agent not being selected for implementation on the environment.

[0010] In some embodiments, the efficiency score indicates a probability of being invoked to generate a suggested action for the environment to resolve the problem.

[0011] The efficiency score may comprise a value between 0 and 1, inclusive. In some embodiments, the efficiency score comprises a value between L and 1, inclusive, where L is a lower limit of the efficiency score, the lower limit being greater than zero and less than 1.

[0012] In some embodiments, based on the efficiency score, invoking the first suggestion agent to generate the suggested action for the environment to solve the problem comprises invoking the first suggestion agent with a probability equal to the efficiency score of the first suggestion agent.

[0013] In some embodiments, invoking the first suggestion agent to generate the suggested action for the environment to resolve the problem is based on a probability value that is independent of the efficiency score of the first suggestion agent.

[0014] In some embodiments, invoking the first suggestion agent to generate the suggested action for the environment to solve the problem is performed using a probability value ε without reference to the efficiency score of the first suggestion agent and a probability value 1-ε based on the efficiency score of the first suggestion agent.

[0015] The efficiency graph may indicate efficiency scores of a plurality of suggestion agents, including the first suggestion agent, in generating suggested actions for the environment to solve the problem.

[0016] The notification may indicate whether the suggested actions generated by the multiple suggestion agents are selected for implementation on the environment, and the method may further include: adjusting the efficiency score of each of the multiple suggestion agents based on whether the suggested actions generated by the multiple suggestion agents are selected for implementation on the environment.

[0017] In some embodiments, the first suggestion agent generates a plurality of suggested actions to resolve the problem, and in response to any one of the plurality of suggested actions being selected for implementation on the environment, the efficiency score of the first suggestion agent is increased.

[0018] The environment may include a computerized system, the information about the environment may include information related to the operation of the computerized system, the intent may include a key performance indicator (KPI) of the computerized system, and the problem may include a failure of the computerized system to meet the KPI.

[0019] The environment may include a communication network, the information about the environment may include information related to the operation of the communication network, the intent may include a key performance indicator (KPI) of the communication network, and the problem may include a failure of the communication network to achieve the KPI. The communication network may be a mobile communication network.

[0020] In some embodiments, the problem includes a first problem, and the efficiency map is associated with a scenario including a plurality of problems including the first problem.

[0021] The method may further include detecting a change in the intent or a change in the problem; and updating the efficiency score in response to the change in the intent or the change in the problem.

[0022] Updating the efficiency score may include increasing or decreasing the efficiency score.

[0023] In some embodiments, updating the efficiency score comprises increasing the efficiency score to a maximum value. In some embodiments, updating the efficiency score comprises increasing the efficiency score by a fixed percentage.

[0024] The method may further include updating the efficiency score after a predetermined period of time has passed. Updating the efficiency score may include increasing or decreasing the efficiency score after the predetermined period of time has passed.

[0025] The method may further include adding one or more edges to the efficiency graph based on a model M that maps effects of performed actions to the problem.

[0026] Some embodiments provide an intent manager for controlling an environment based on an intent for the environment. The intent manager includes a reasoner that receives information about the environment and, based on the information about the environment, identifies a problem in the environment that needs to be solved to achieve the intent. The reasoner obtains an efficiency graph associated with the problem, the efficiency graph indicating an efficiency score of a first suggestion agent in generating a suggested action for the environment to solve the problem. Based on the efficiency score, the reasoner invokes the first suggestion agent to generate a suggested action for the environment to solve the problem.

[0027] The intent manager further includes an evaluation agent that determines whether the suggested action generated by the first suggestion agent is selected for implementation on the environment and notifies the reasoner whether the suggested action generated by the first suggestion agent is selected for implementation on the environment. The reasoner adjusts the efficiency score of the first suggestion agent based on whether the suggested action generated by the first suggestion agent is selected for implementation on the environment.

[0028] The intent manager may also include a prediction agent that estimates the impact of the suggested action generated by the first suggestion agent on the environment, wherein the evaluation agent determines whether the suggested action generated by the first suggestion agent should be selected for implementation on the environment based on the estimated impact of the suggested action on the environment.

[0029] The intent manager may also include an actuator agent that, in response to the evaluation agent determining that the suggested action should be implemented on the environment, performs the suggested action on the environment.

[0030] Some embodiments described herein can help achieve a more efficient system by enabling the system for controlling an environment to generate and select better suggestions. Additionally, some embodiments can improve the overall efficiency of the system in terms of scalability by reducing the number of potential suggestions to be considered. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 A system including an intent manager for controlling an environment using zero-touch control is shown.

[0032] Figure 2 Units of the intent manager for controlling an environment using zero-touch control are shown in more detail.

[0033] Figure 3 An example of a mobile communication network is shown which may be an environment controlled by an intent manager.

[0034] Figure 4 The architecture of an intent manager according to some embodiments is shown.

[0035] Figure 5 is an example of a graph showing the relationship between multiple agents and problems.

[0036] Figure 6 Examples of efficiency diagrams are shown for two possible cases.

[0037] Figure 7 An efficiency graph is shown as stored in a knowledge base of an intent manager framework according to some embodiments.

[0038] Figure 8 The operation of various units of the intent manager according to some embodiments is shown.

[0039] Figure 9 An example of applying the embodiments described herein to a network use case in which an intent manager is deployed is shown.

[0040] Figure 10The operation of the intent manager according to some embodiments is shown.

[0041] Figure 11A is a block diagram illustrating elements of an intent manager according to some embodiments.

[0042] Figure 11B Various functional modules are shown that are stored in the memory of the intent manager according to some embodiments. DETAILED DESCRIPTION

[0043] As mentioned above, a problem arises when a measured KPI doesn't meet its target. This problem needs to be addressed, and the first step in achieving this goal is to recommend one or more actions. These actions, if taken, could bring the KPI closer to its target. Multiple suggestion agents can generate recommendations for actions that should be taken to achieve the desired intent. However, to determine the recommended action, the intent manager must invoke the correct suggestion agent.

[0044] One approach is for each suggestion agent 34b to register itself with the intent manager 10. The suggestion agent 34b identifies problems that it can suggest actions to solve. For example, assume that the measurement value of the latency KPI for a particular service KPI is 30 milliseconds of latency. Further assume that the target latency value is a maximum of 20 milliseconds. The problem to be solved is that the latency needs to be improved by at least 10 milliseconds. In this example, there can be a suggestion agent that registers itself with information that it can suggest actions for the latency problem. When the latency problem suggestion agent is called, it can perform a root cause analysis and, based on the results, suggest actions to solve the latency problem, such as increasing the user-plane priority of users involved in the service, decreasing the user-plane priority of users of another service, increasing the data center computing capacity allocated to the service, and the like.

[0045] In a second approach, the suggestion agent 34b will register the types of actions it can suggest, rather than registering the ability to solve a specific problem. Following the example above, multiple suggestion agents 34b will register with the intent manager 10. For example, one agent may register that it can suggest the action "lower the user plane priority of certain services." A second suggestion agent 34b may register that it can suggest the action "increase the user plane priority of services." A third suggestion agent 34b may register that it can suggest the action "increase the data center computing capacity of services."

[0046] In the second approach, each suggestion agent 34b can still perform some root cause analysis, such as investigating the severity of the problem to determine how much the user plane priority should be increased, or for which service the user plane priority should be increased. Nevertheless, the scope of the root cause analysis is smaller than that of the first approach. And, more importantly, additional logical reasoning rules are required to tell the system that a specific agent needs to be called when a latency problem occurs. In this example, there may be multiple such rules, such as one rule saying, "When a latency problem occurs, call a suggestion agent that reduces the user plane priority," or another rule saying, "When a latency problem occurs, call a suggestion agent that increases the data center capacity," and so on.

[0047] Note that these two approaches are not mutually exclusive and they can be used simultaneously in the same intent manager 10.

[0048] In both approaches, the intent manager 10 does not learn over time. Both approaches rely on decision rules, which are provided by the suggestion agent 34b, as in the first approach, or by another entity outside of the suggestion agent 34b, as in the second approach. In this regard, the design is quite static because the rules do not change automatically, and therefore, the intent manager 10 does not learn over time. This leads to inefficiencies. For example, in the second approach, many actions are suggested, but most of these actions will be discarded by the evaluation agent 34c. Because the system cannot learn, inaccurate actions for a given problem may be repeatedly suggested. Another problem is missing solutions. With static rules, potential solutions to a given problem may not be covered in the rules because the human designer did not consider this possibility. Such missing solutions (manifested in the form of missing action suggestions) will never be suggested and, therefore, will never be evaluated and executed.

[0049] Certain aspects of the present disclosure and its embodiments may provide solutions to these and other challenges. Some embodiments described herein provide a more dynamic approach in which the intent manager 10 learns over time which action suggestions are best for a given problem. According to some embodiments, the decision rules for which suggestion agent 34b to invoke are dynamic and can be added, changed, or deleted based on insights gained by the intent manager 10 over time.

[0050] Therefore, for a problem that occurs in the controlled environment 20, the intention manager 10 can call all available suggestion agents 34b registered for solving the problem. That is, the intention manager 10 initially has no idea which suggestion agents 34b are best suited to solve the relevant problem.

[0051] Each suggestion agent 34b generates one or more suggested actions, and the intent manager 10 performs prediction / evaluation operations to decide what the expected outcome of each suggested action is.

[0052] The intent manager 10 monitors which suggested actions are accepted / rejected by the evaluation agent 34c. Accepted / rejected suggestions are traced back to the knowledge base for invoking the suggestion agent 34b. Rejected suggestions penalize the suggestion agent 34b that made the suggestion, while accepted suggestions improve the efficiency of the suggestion agent 34b that made the suggestion.

[0053] When the system changes in some way, e.g., providing new intents, adding new actuator agents, etc., the system may slowly forget the previously learned efficiencies.

[0054] Certain embodiments can provide one or more of the following technical advantages. Some embodiments described herein can contribute to a more efficient system by enabling a system for controlling an environment to generate and select better recommendations. Additionally, some embodiments can improve the overall efficiency of the system in terms of scalability by reducing the number of potential recommendations to consider.

[0055] Some embodiments contemplated herein will now be described more fully with reference to the accompanying drawings.The embodiments are provided by way of example to convey the scope of the subject matter to those skilled in the art.

[0056] As mentioned above, Figure 2 The illustrated environment 20 may include a computerized system, such as a communications network or a portion thereof. Figure 3 An example of a mobile communications network 50 is shown. Specifically, mobile communications network 50 comprises a radio access network (including gNodeBs / gNBs), a core network containing various network functions (such as the User Plane Function (UPF), Policy Control Function (PCF), Session Management Function (SMF), Application Function (AF), and so on), as well as applications. Note that all of these functions can be distributed across multiple physical sites. For example, multiple UPF instances can exist at different sites. Similarly, these applications can be distributed as multiple instances at local edge sites (near the UPF) or at national hub sites. Figure 3 Examples of actions that can be taken on the network are also shown. These actions (shown in Application Programming Interface (API) definitions) would be actions that the actuator agent can take on the mobile communication network 50.

[0057] Figure 4The architecture of the intent manager 10 according to some embodiments is shown. Specifically, the intent manager 10 includes an intent manager framework 30, which manages the operation of multiple agents, including one or more data anchor agents 34a, suggestion agents 34b, prediction agents 34e, evaluation agents 34c, and actuator agents 34d. The intent manager framework 30 includes a knowledge base 32, which stores KPIs, questions, network status, and one or more efficiency graphs, as discussed in more detail below. The intent manager framework 30 also includes a reasoner 36, which contains logic for controlling the actions of the agents 34a-34e based on the information stored in the knowledge base 32 and the intent received from the operator 12.

[0058] The intent manager framework 30 manages the operation of the intent manager 10 to implement the intent provided by the operator 12. The data anchor agent 34a receives raw data from the environment 20 and prepares the raw data for use by the suggestion agent. The raw data is also provided to the reasoner 36 via the knowledge base 32 of the intent manager framework 30, and the reasoner 36 analyzes the data to identify problems that need to be solved.

[0059] When a target KPI is not met, the reasoner 36 creates a knowledge object called a problem. For example, when the environment is a communication network, the information about the environment may include information related to the operation of the communication network, the intent may be a KPI for the communication network, and the problem may be that the communication network failed to meet the KPI.

[0060] The suggestion agent 34b identifies the problem and assigns itself to generate a solution. The suggestion agent 34b can be diversified as needed and can implement many different procedures (e.g., if-then-else rules, reinforcement learning, optimization, and a combination of root cause analysis and rules) to solve the problem represented by the problem.

[0061] After one or more suggestion agents 34b recommend an action, a prediction agent 34e estimates the impact of the suggestion on environment 20 before the suggested action is potentially executed. Conflicts are then analyzed against target KPIs. For example, a suggestion might solve one problem but create or exacerbate another. Therefore, after the prediction agent 34e has predicted the impact of the suggested action on environment 20, the evaluation agent 34c resolves any conflicts identified during the prediction process. Finally, the approved action is executed by the actuator agent 34d.

[0062] According to some embodiments, the intent manager framework 30 learns which suggestion agents 34b to invoke based on which problems exist. Figure 4 The arrow between the suggestion agent 34b and the intent manager framework in indicates the communication required for such a "smart" call.

[0063] In the previous approach, when a given problem is solved, all suggestion agents 34b registered to solve that problem will be called. All suggestions generated by the suggestion agents 34b will need to be evaluated. This can lead to some inefficiencies if the suggestion agents 34b generate suggestions that are unlikely to be selected.

[0064] In the intent manager 10, multiple suggestion agents can be registered to solve a specific problem. For example, Figure 5 is an example of a graph showing the relationship between multiple agents (A, B, C) and problems (X, Y). Figure 5 In the example above, agent A claims that it can solve problem X and problem Y, while agents B and C can solve problem Y. In previous approaches, when a problem, such as an unmet KPI, is identified, all suggestion agents registered to solve that problem are invoked. In contrast, some embodiments described herein provide a more dynamic and intelligent way to invoke suggestion agents.

[0065] In the following discussion, a "situation" is defined as a specific set of problems that exist in a system. For example, Figure 5 , in the first case (Case 1), both Problem X and Problem Y exist. In the second, different case (Case 2), only Problem X exists.

[0066] According to some embodiments, the intent manager 10 tracks the effectiveness of all suggestion agents 34b in resolving a particular situation and invokes agents based on their effectiveness. In this context, when a suggestion agent 34b is invoked, it is requested by the intent manager framework 30 to generate a suggestion for an action to be taken to resolve an identified problem. The effectiveness of the suggestion agent 34b with respect to a particular situation is variable and is adjusted in each iteration based on whether a suggestion generated by the suggestion agent 34b is selected by the evaluation agent 34c for application to the environment 20.

[0067] The effectiveness of the suggestion agent 34b may be represented on a so-called effectiveness graph which, for a given situation, shows both the relationship between the suggestion agent 34b and the problem, and the effectiveness of the suggestion agent 34b in solving the problem presented in the situation.

[0068] Figure 6 Example efficiency graphs 40a, 40b are shown for two possible scenarios (Scenario 1 and Scenario 2). The roots of the efficiency graphs 40a, 40b represent the problems present in the scenario, and the leaves of the efficiency graphs correspond to the recommended agents 34b that can be invoked to resolve the problems present in the scenario. For example, in Figure 6In [1], when both problem X and problem Y exist, there is case 1, and when only problem Y exists, there is case 2. Only the proposed agent A is registered as being able to solve problem X, while the proposed agents A, B, and C are all registered as being able to solve problem Y.

[0069] An efficiency score (Eff_score) is associated with each leaf in Figures 40a and 40b. The efficiency score can be a value in the range [L, 1]. In some embodiments, the efficiency score represents the probability of invoking the recommended agent for that situation. L, between 0 and 1, is the minimum allowable value for efficiency. L can be defined at design time or adjusted at runtime.

[0070] In some embodiments, the efficiency score may be used as an absolute probability for invoking the agent. Alternatively, the intent manager 10 may employ an ε-greedy approach to invoking and invoke a suggestion with a certain probability Y independent of the suggestion's efficiency score.

[0071] ε-Greedy is a common method in reinforcement learning for selecting actions to balance exploration and exploitation. It involves selecting a random action with probability ε (a small value, typically 0.1 or 0.01) and selecting the action with the highest expected reward with probability 1-ε. This allows the agent to try new actions and explore its environment while leveraging its acquired knowledge about which actions tend to lead to high rewards.

[0072] In one example of the ε-greedy approach, there is a 5% probability that the agent is always called regardless of the agent's efficiency score, and a 95% probability that the call depends on the efficiency score (as described above).

[0073] As in Figure 5 In the example shown, when all efficiency scores are equal to 1, the system behaves similarly to the baseline system. That is, for each problem, all suggestion agents 34b registered to solve that problem are invoked. According to some embodiments, the system always starts in this state without knowledge. However, as suggestions are proposed, selected, and applied to the environment 20, the efficiency scores are adjusted over time.

[0074] Figure 7 The efficiency graphs 40a-40n are shown to be stored in the knowledge base 32 of the intent manager framework 30. The efficiency graphs 40a-40n can be updated at runtime when a new intent (with a new expectation) is registered. This triggers the creation of a new efficiency graph in the knowledge base.

[0075] The efficiency graphs 40a, 40b may also be updated when an advice agent is registered / unregistered in the intent manager 10. This triggers an update of the leaves of the efficiency graph that already exist in the knowledge base.

[0076] The reasoner 36 of the intent manager framework 30 acts as an inference mechanism for invoking the suggestion agent 34 b based on information contained in the efficiency maps 40 a , 40 b stored in the knowledge base 32 .

[0077] As an example, suppose there is Figure 5 Consider situation 1, where problems X and Y exist. In this example, each suggestion agent 34b recommends a single action. For a new problem, the reasoner 36 invokes all available suggestion agents "registered" to solve the problem with a probability based on their efficiency scores. Therefore, agents A, B, and C will be invoked because their efficiency scores are 1 (i.e., probability of invocation = 1). Each suggestion agent A, B, and C generates one or more suggested actions. The suggested actions are analyzed by the prediction agent 34e and the evaluation agent 34c. The intent management framework 30 monitors which suggested actions are accepted / rejected by the evaluation agent 34c and traces these actions back to the knowledge base 32 used to invoke the suggestion agent 34b. For example, action 2 from agent 2 may be accepted, while actions 1 and 3 (from agents 1 and 3) may be rejected (i.e., discarded). When a suggestion is rejected, the suggestion agent 34b that made the suggestion is penalized, and its efficiency score relative to the specific situation being handled decreases. When a suggestion is accepted, the efficiency score of the suggestion agent 34b that made the accepted suggestion increases, up to a maximum of 1.

[0078] As a result of these actions, when the situation arises again, agent 2 will be called with a higher probability than the other agents (because it was approved).

[0079] Abstracting from the above specific examples, Figure 8 The sequence diagram illustrates operations according to some embodiments, wherein invocation of an agent is based on an efficiency score of the agent, and wherein an efficiency map can be dynamically updated.

[0080] See also Figure 8 In step 1, the reasoner 36 uses data obtained from the environment and information stored in the knowledge base 32 to evaluate the environment KPIs. In step 2, the reasoner 36 generates the problem to be solved (i.e., the unmet expectations). In step 3, the current problem is used to define the situation, and the efficiency graph 40a representing the situation is retrieved from the knowledge base 32 (step 4).

[0081] In step 5, the reasoner 36 invokes one or more suggested agents 34b based on the information in the efficiency map 40a. That is, the reasoner 36 determines whether to invoke a particular suggested agent 34b based on an invocation probability derived from the information stored in the efficiency map 40a. For example, in some embodiments, the suggested agent 34b may be invoked with a probability equal to the efficiency score of the suggested agent 34b. In other embodiments, there may be a fixed probability (ε) that the suggested agent 34b is invoked, and a probability (1-ε) that the suggested agent 34b is invoked based on the efficiency score of the suggested agent 34b (i.e., a probability ε that the suggested agent 34b will be invoked, and a probability 1-ε that the suggested agent 34b may or may not be invoked based on the efficiency score of the suggested agent 34b).

[0082] In step 6, suggestion agent 34b suggests actions that will be evaluated by prediction agent 34e in step 7. Prediction agent 34e sends the actions and their predicted effects to evaluation agent 34c in step 8. In step 9, evaluation agent 34c selects an action to be performed.

[0083] In step 10, the evaluation agent 34c notifies the reasoner 36 of the selected action. In step 11, the evaluation agent 34c sends the action to the actuator agent for execution.

[0084] In step 12, the reasoner 36 updates the efficiency map 40a based on which actions were selected. As described above, if an action is selected, the efficiency score of the suggestion agent 34b that suggested the action increases, and if an action is not selected, the efficiency score of the suggestion agent 34b that suggested the action decreases.

[0085] In the case where an agent suggests multiple actions, if at least one of the suggested actions is accepted after evaluation, the agent's efficiency score increases. Optionally, the intent manager 10 can send feedback to the suggestion agent 34b that was rejected, asking them not to suggest the rejected action for this situation in the future.

[0086] In some embodiments, the efficiency score can be updated after evaluation by the evaluation agent 34c. In some embodiments, a fixed update can be implemented. For example, a reward / penalty constant K (e.g., K = 0.1) can be defined at design time. If the action recommended by the suggestion agent 34b is approved, the new efficiency score of the suggestion agent 34b is increased by K; otherwise, the efficiency score is decreased by K.

[0087] In some embodiments, the efficiency score can be updated dynamically. For example, in the intent manager 10, each intent can have an associated penalty formula. The penalty can be a function of the penalties for all intents, such as the sum of all intent penalties. In a dynamically updated embodiment, the reward / penalty is proportional to the incremental intent manager penalty after prediction / evaluation, where the incremental penalty is the current intent manager penalty minus the predicted intent manager penalty for the suggested action.

[0088] In some embodiments, when something changes within intent manager 10, such as when a new intent is provided, intent manager 10 can update the efficiency score for each efficiency graph. This is to allow for better learning across existing and new situations and to prevent learning drift. By doing this, the intent manager can slowly forget what it has learned as dynamic events occur. For example, the efficiency score can be reset to or increased to a value of 1.0 (indicating the highest probability of being invoked).

[0089] In some embodiments, a constant U can be defined at design time to smooth out variations in the efficiency scores of the suggestion agent 34b that can be invoked for a given situation. For example, if U=0.2, all efficiency scores can be increased by 20% when a new intent is received, a new expectation is provided, or a predetermined time period has passed.

[0090] This update may also be triggered upon notification from another module responsible for identifying the level of dynamics in the network in a more accurate manner.

[0091] In some embodiments, when a new situation arises, the intent manager can employ transfer learning rather than initializing a new efficiency graph with an efficiency score of 1 due to new expectations being registered. In this case, the agent's efficiency score for the new situation can be initialized with the efficiency score of the most similar situation already in the system. The most similar situation can be defined as the situation that has the most common problems with the new situation.

[0092] Another aspect of dynamics is the passage of time. For example, network traffic conditions are often very different during the day and at night. Problems that occur during peak hours may not occur during off-peak hours. To accommodate this dynamic nature, the system can simply update the efficiency score upward (i.e., toward the highest probability) after a certain period of time has passed.

[0093] Another aspect of dynamism involves creating new edges between agents and problems. By creating new edges in the efficiency graph, the system learns new rules and increases the number of suggestions that can solve the problem. Without the efficiency graph, this approach could hinder the system by increasing the number of recommended suggestions. However, the efficiency graph makes this approach more scalable because it prunes inefficient recommendations.

[0094] Adding new edges aims to identify actions that, when executed, affect the KPIs expressed in the question. One possible implementation is based on supervised learning. In supervised learning, the system attempts to learn a model M that maps the effects of executed actions to the KPIs expressed in expectations. The impact of an action is defined by the difference between the KPI values before and after execution. M is defined as follows: M(O, S0, S1, C0) = (C1, A1,), where: O: represents the parameters of model M; S0: represents the KPI value before the action is executed; S1: represents the KPI value after the action is executed; C0: indicates the current network configuration; C1: Recommended action (or new network configuration); A1: The agent that suggested the action.

[0095] Once M is available, M can be queried to add new edges to the efficiency graph, such as Figure 6 As shown in Figure 1, for example, in Case 1, the system queries model M for the current state (S0), the target state (S1) for fixing problem X, and the current network configuration (C0). Based on the output of model M, if it is determined that the actions agent B can take can solve problem X, a new edge can be added between agent B and problem X.

[0096] Other mechanisms and formulas can be used to add new edges to the efficiency graph. Since the efficiency graph can be expanded, expanding the number of suggestions is feasible because the efficiency graph ensures that the number of suggestions remains manageable.

[0097] Figure 9 An example of applying the embodiments described herein to a network use case in which the intent manager 10 is deployed is shown. Specifically, Figure 9 In the example shown, a URLLC intent is submitted to the intent manager 10. Two non-functional expectations associated with URLLC applications are: latency < 50ms for all users of the service, and packet loss less than 1% for all users of the service.

[0098] Since this is the only intent in the system, all of its expectations are now met. However, suppose a new intent is inserted into the system that requires that the QoE for 80% of users of the video service must be greater than 4.0. As is well known, video traffic is very intensive on network resources, and therefore, it causes URLLC applications to perform poorly. Specifically, while the expectations of the video service are fully met, some URLCC expectations may now be unmet. By applying the systems / methods described herein, the best / correct suggestion agent 34b can be invoked based on the severity of the situation for the URLLC service.

[0099] Figure 9 Examples of efficiency graphs are shown for two different cases, i.e., where only packet loss is not achieved for the URLLC service. Figure 1 , and the efficiency where both packet loss and latency issues exist Figure 2 For both cases, two suggestion agents 34bv may be invoked, including a suggestion agent 34b that increases the priority of the application and a suggestion agent 34b that moves the context of the application to a new site.

[0100] In the baseline approach, both suggestion agents 34b will be invoked to try to resolve the issue, regardless of which situation exists. However, this will result in inefficiency because all action suggestions will need to be predicted and evaluated, which requires more computation and, therefore, consumes more energy on the hardware on which the intent manager 10 is running.

[0101] Furthermore, some recommendations may take a long time to evaluate, risking hindering the closed loop of monitoring, recommendation, prediction, and actuation, and thus increasing the closed loop resolution time.

[0102] According to some embodiments, the intention manager 10 has already faced the current situation and has stored and updated the relevant efficiency map 40a in the knowledge base 32.

[0103] When faced with situation 1 (i.e., packet loss problem, such as Figure 9 ), the suggestion agent 34b that suggests increasing the application priority may be invoked with a higher probability than the suggestion agent 34b that suggests moving the application context.

[0104] In practice, in this scenario, it might be found that moving the application context is efficient but very costly for the intent manager (e.g., a higher penalty). On the other hand, an action to increase the priority might still be efficient while requiring less cost (e.g., moving the application in a new site might require instantiating a new virtual machine to which the application context is moved). Therefore, in this case, there is a trade-off for efficiency. Figure 1 , the following holds: - Mobile apps: eff_score = 0.2 - Improve priority: eff_score=0.98

[0105] When faced with situation 2 (i.e., packet loss and delay issues, such as Figure 9 As shown on the right side of the figure, the suggestion agent 34b that suggests a mobile application context may be invoked with a higher probability than the suggestion agent 34b that suggests increasing the application priority. In other words, the opposite of the previous case may occur. Increasing the priority of the URLLC application may not be sufficient to resolve these issues; a full mobile application context may be required to address both latency and packet loss. Therefore, in this case, the following holds true: - Mobile app: eff_score = 0.99 - Increase priority: eff_score=0.5

[0106] Some embodiments can be applied to other use cases, as the concepts of "problems" or "issues" and the "agents" that solve them can be similarly instantiated for different use cases. For example, the environment can include a real-world environment, such as a technical system that includes sensors, actuators, processors, and other components. For example, the technical system can include a communication network, where the "agent" can be an electronic agent or a software agent, etc. The communication network includes actuators, such as base stations, that can be controlled to affect network operation.

[0107] In some embodiments, the environment may be a smart factory including smart mechanical systems including processing circuits and actuators, such as robots, conveyor belts, etc., controlled by a processing unit.

[0108] In smart factories, robots with specific capabilities can register with the robot orchestrator to solve problems or requests (for example, tasks like moving packages or packing). When a request is set, the intent manager learns which robot is best suited for deployment. For simple cases, a general-purpose robot can be deployed to solve the problem. On the other hand, complex situations require highly specialized robots that are costly to operate.

[0109] By applying the embodiments described herein to the context of a smart factory use case, a robot manager can learn over time which group of robots to call upon to solve the task at hand. This allows the orchestrator to, for example, determine whether it should call upon one specialized robot or multiple general-purpose robots.

[0110] A similar approach can be applied to learn where to dispatch support calls (i.e., to which call center operator) based on the customer's problem at hand. Different operators can be evaluated based on their problem-solving abilities. When a customer at the call center registers to address a specific issue, the dispatch call system automatically redirects the call to an operator who is expected to perform better for that situation.

[0111] Figure 10 FIG2 illustrates the operation of the intent manager 10 to control an environment based on intent for the environment 20 according to some embodiments. Figure 4 and 10 The method includes receiving information about an environment (1002). For example, the information may be received from the environment via a data anchor agent 34a, which receives raw data from the environment, processes the data, and stores the data in a knowledge base 32 in an intent manager framework 30 of the intent manager 10. A reasoner 36 in the intent manager framework 30 may retrieve data from the knowledge base 32.

[0112] Based on the information about the environment, the reasoner 36 identifies problems in the environment that need to be solved to achieve the intent (block 1004 ).

[0113] After identifying a problem, the reasoner 36 obtains an efficiency graph associated with the problem (block 1006). For example, the efficiency graph may be retrieved from the knowledge base 32. The efficiency graph indicates the efficiency scores of the plurality of suggestion agents 34b, including the first suggestion agent 34b, in generating suggested actions for the environment to solve the problem. In some embodiments, the efficiency graph may be associated with a scenario that includes multiple problems and associates a suggestion agent 34b with a problem that the suggestion agent can solve.

[0114] Based on the efficiency score of the first suggestion agent, the reasoner 36 invokes the first suggestion agent 34b to generate a suggested action for the environment to solve the problem (block 1008). Specifically, the reasoner 36 may decide to invoke the first suggestion agent 34b based on a probability function that uses the efficiency score of the first suggestion agent 34b as an input parameter, as described above.

[0115] The intent manager 10 then determines whether the suggested action generated by the first suggestion agent is selected for implementation on the environment (block 1010). This determination can be made by generating a prediction of the effect the suggested action will have on the environment 20 and evaluating the suggested action based on the prediction.

[0116] The reasoner 36 then adjusts the efficiency score of the first suggestion agent based on whether the suggested action generated by the first suggestion agent is selected for implementation on the environment (block 1012 ).

[0117] Then, at block 1014, the reasoner 36 determines whether the problem has been resolved. If so, the process ends. Otherwise, the operation returns to block 1006, where the reasoner 36 obtains the modified efficiency graph and repeats the process of invoking the suggestion agent based on the modified efficiency graph.

[0118] Figure 11A is a block diagram illustrating elements of an intent manager 10 according to some embodiments. Intent manager 10 can be provided by, for example, a device in the cloud running software on cloud computing hardware, or by a software function / service that manages or controls a communication network. That is, the device can be implemented as part of a communication system or as a separate function / service hosted in the cloud on a device. The device can also be provided as standalone software for managing a communication network, and can be part of a deployment that includes virtual or cloud-based network functions (VNFs or CNFs) or even physical network functions (PNFs). The cloud can be public, private (e.g., on-premises or hosted), or a hybrid cloud.

[0119] As shown, the device may include transceiver circuitry 1101 (e.g., RF transceiver circuitry) comprising a transmitter and a receiver configured to provide uplink and downlink radio communications with a device (e.g., a controller for automatically performing an actuation). The device may also include network interface circuitry 1108 (also referred to as a network interface) configured to provide communications with other devices (e.g., a controller for automatically performing an actuation). The device may also include processing circuitry 1103 (also referred to as a processor) coupled to the transceiver circuitry and memory circuitry 1105 (also referred to as a memory) coupled to the processing circuitry.

[0120] As described herein, the operations of the device may be performed by processing circuitry 1103, network interface 1108, and / or transceiver 1101. For example, processing circuitry 1103 may control intent manager 10 to perform operations according to the embodiments disclosed herein. Processing circuitry 1103 may also control transceiver 1101 to transmit downlink communications to one or more devices via a radio interface via transceiver 1101 and / or to receive uplink communications from one or more devices via transceiver 1101 via a radio interface. Similarly, processing circuitry 1103 may control network interface 1108 to transmit communications to one or more devices via network interface 1108 and / or to receive communications from one or more devices via the network interface. Furthermore, modules may be stored in memory 1105, and these modules may provide instructions such that, when processing circuitry 1103 executes the instructions of the modules, processing circuitry 1103 performs corresponding operations (e.g., the operations discussed below with respect to example embodiments related to the device). According to some embodiments, intent manager 10 and / or its units / functions may be embodied as one or more virtual devices and / or one or more virtual machines.

[0121] According to some other embodiments, the device may be implemented without a transceiver. In such embodiments, transmissions to the wireless device may be initiated by the intent manager 10, thereby providing transmissions to the wireless device via a device including a transceiver (e.g., via a base station). According to embodiments in which the device includes a transceiver, initiating the transmission may include sending via the transceiver.

[0122] Figure 11B Various functional modules are shown as being stored in the memory 1105 of the intent manager 10. Specifically, the memory 1105 may include a reasoner module 1112 that implements the reasoner 36 of the intent manager 10. The memory 1105 may also include a data anchor module 1114 that implements the data anchor agent 34a, a suggestion module 1116 that implements the suggestion agent 34b, an evaluation module 1118 that implements the evaluation agent 34c, an actuator control module 1120 that implements the actuator agent 34d, and / or a prediction module 1122 that implements the prediction agent 34e. It should be understood that in various embodiments, one or more of the data anchor agent 34a, the suggestion agent 34b, the evaluation agent 34c, the actuator agent 34d, and / or the prediction agent 34e may be implemented separately from the intent manager 10.

[0123] Other modules may also be provided in the memory 1205 to implement the above operations.

[0124] While the computing devices (e.g., UEs, network nodes, hosts) described herein may include combinations of the hardware components shown, other embodiments may include computing devices with different combinations of components. It should be understood that these computing devices may include any suitable combination of hardware and / or software necessary to perform the tasks, features, functions, and methods disclosed herein. The determinations, calculations, acquisitions, or similar operations described herein may be performed by processing circuitry, which may, for example, process information by converting the acquired information into other information, comparing the acquired information or the converted information with information stored in the network node, and / or performing one or more operations based on the acquired information or the converted information, and making determinations as a result of such processing. Furthermore, while components may be described as a single block within a larger block, or nested within multiple blocks, in practice, a computing device may include multiple distinct physical components that make up the single illustrated component, and functionality may be divided between separate components. For example, a communication interface may be configured to include any of the components described herein, and / or functionality of a component may be divided between the processing circuitry and the communication interface. In another example, the non-computationally intensive functionality of any such component may be implemented in software or firmware, while the computationally intensive functionality may be implemented in hardware.

[0125] In certain embodiments, some or all of the functionality described herein may be provided by processing circuitry executing instructions stored in a memory, which in certain embodiments may be a computer program product in the form of a non-transitory computer-readable storage medium. In alternative embodiments, some or all of the functionality may be provided by processing circuitry without, for example, hard-wired execution of instructions stored on a separate or separate device-readable storage medium. In any of these certain embodiments, the processing circuitry may be configured to perform the described functionality regardless of whether or not instructions are executed on a non-transitory computer-readable storage medium. The benefits provided by such functionality are not limited to individual processing circuitry or other components of a computing device, but rather are generally benefited by the entire computing device, as well as by end users and wireless networks.

Claims

1. A method performed by an intent manager (10) for controlling an environment based on an intent for the environment, the method comprising: receiving (1102) information about the environment; Based on the information about the environment, identifying ( 1104 ) a problem in the environment that needs to be solved to achieve the intent; obtaining ( 1106 ) an efficiency graph associated with the problem, wherein the efficiency graph indicates an efficiency score of the first suggestion agent in generating suggested actions for the environment to solve the problem; Based on the efficiency score, calling ( 1108 ) the first suggestion agent to generate a suggested action for the environment to resolve the problem; determining ( 1110 ) whether the suggested action generated by the first suggestion agent is selected for implementation on the environment; and The efficiency score of the first suggestion agent is adjusted ( 1112 ) based on whether the suggested action generated by the first suggestion agent is selected for implementation on the environment.

2. The method according to claim 1, wherein Adjusting the efficiency score of the first suggestion agent includes increasing the efficiency score of the first suggestion agent in response to the suggested action generated by the first suggestion agent being selected for implementation on the environment, and decreasing the efficiency score of the first suggestion agent in response to the suggested action generated by the first suggestion agent not being selected for implementation on the environment.

3. The method according to claim 1 or 2, wherein: The efficiency score indicates a probability of being invoked to generate a recommended action for the environment to resolve the problem.

4. The method according to claim 3, wherein: The efficiency score comprises a value between 0 and 1, inclusive.

5. The method according to claim 3, wherein The efficiency score comprises a value between L and 1, inclusive, where L is a lower limit of the efficiency score, the lower limit being greater than zero and less than 1.

6. A method according to any preceding claim, wherein: Based on the efficiency score, invoking the first suggestion agent to generate the suggested action for the environment to solve the problem includes invoking the first suggestion agent with a probability equal to the efficiency score of the first suggestion agent.

7. The method according to claim 6, wherein: Invoking the first suggestion agent to generate the suggested action for the environment to resolve the problem is based on a probability value that is independent of the efficiency score of the first suggestion agent.

8. The method according to claim 7, wherein: Invoking the first suggestion agent to generate the suggested action for the environment to solve the problem is performed using a probability value ε without reference to the efficiency score of the first suggestion agent and a probability value 1-ε based on the efficiency score of the first suggestion agent.

9. A method according to any preceding claim, wherein: The efficiency graph indicates efficiency scores of a plurality of suggestion agents, including the first suggestion agent, in generating suggested actions for the environment to solve the problem.

10. The method according to claim 9, wherein: The notification indicates whether the suggested actions generated by the plurality of suggestion agents are selected for implementation on the environment, the method further comprising: The efficiency score of each of the plurality of suggestion agents is adjusted based on whether the suggested actions generated by the plurality of suggestion agents are selected for implementation on the environment.

11. A method according to any preceding claim, wherein: The first suggestion agent generates a plurality of suggested actions to solve the problem, and wherein the efficiency score of the first suggestion agent is increased in response to any one of the plurality of suggested actions being selected for implementation on the environment.

12. A method according to any preceding claim, wherein: The environment includes a computerized system, the information about the environment includes information related to operation of the computerized system, the intent includes a key performance indicator (KPI) of the computerized system, and the problem includes a failure of the computerized system to achieve the KPI.

13. The method according to claim 12, wherein: The environment comprises a communication network, the information about the environment comprises information related to operation of the communication network, the intent comprises a key performance indicator (KPI) of the communication network, and the problem comprises a failure of the communication network to meet the KPI.

14. A method according to any preceding claim, wherein: The problems include a first problem, and wherein the efficiency map is associated with a situation including a plurality of problems including the first problem.

15. The method according to any preceding claim, further comprising: detecting a change in the intent or a change in the question; as well as The efficiency score is updated in response to the change in the intent or the change in the problem.

16. The method according to claim 15, wherein Updating the efficiency score includes increasing the efficiency score.

17. The method according to claim 16, wherein: Updating the efficiency score includes increasing the efficiency score to a maximum value.

18. The method according to claim 16, wherein Updating the efficiency score includes increasing the efficiency score by a fixed percentage.

19. The method according to claim 15, wherein Updating the efficiency score includes decreasing the efficiency score.

20. The method of any preceding claim, further comprising: After a predetermined period of time has passed, the efficiency score is updated.

21. The method according to claim 20, wherein Updating the efficiency score includes increasing the efficiency score after the predetermined time period has elapsed.

22. The method according to claim 20, wherein Updating the efficiency score includes decreasing the efficiency score after the predetermined time period has elapsed.

23. The method of any preceding claim, further comprising: One or more edges are added to the efficiency graph based on a model M that maps the effects of the performed actions to the problem.

24. An intention manager (10) for controlling an environment based on an intention for the environment, the intention manager comprising: a reasoner (36) that receives information about the environment and, based on the information about the environment, identifies a problem in the environment that needs to be solved to achieve the intent, wherein the reasoner obtains an efficiency graph associated with the problem, wherein the efficiency graph indicates an efficiency score of a first suggestion agent (34b) in generating a suggested action for the environment to solve the problem, and wherein, based on the efficiency score, the reasoner invokes the first suggestion agent to generate a suggested action for the environment to solve the problem; and an evaluation agent (34c) that determines whether the suggested action generated by the first suggestion agent is selected for implementation on the environment and notifies the reasoner whether the suggested action generated by the first suggestion agent is selected for implementation on the environment; wherein the reasoner adjusts the efficiency score of the first suggestion agent based on whether the suggested action generated by the first suggestion agent is selected for implementation on the environment.

25. The intent manager of claim 23, further comprising: a prediction agent (34e) that estimates an impact of the suggested action generated by the first suggestion agent on the environment, wherein the evaluation agent determines whether the suggested action generated by the first suggestion agent should be selected for implementation on the environment based on the estimated impact of the suggested action on the environment.

26. The intent manager of claim 24 or 25, further comprising: An actuator agent (34d) that, in response to the evaluation agent determining that the recommended action should be implemented on the environment, performs the recommended action on the environment.

27. An intent manager according to any one of claims 24 to 26, wherein: The reasoner adjusts the efficiency score of the first suggestion agent by increasing the efficiency score of the first suggestion agent in response to the suggested action generated by the first suggestion agent being selected for implementation on the environment, and decreasing the efficiency score of the first suggestion agent in response to the suggested action generated by the first suggestion agent not being selected for implementation on the environment.

28. An intent manager according to any one of claims 24 to 27, wherein: The efficiency score indicates a probability of being invoked to generate a recommended action for the environment to resolve the problem.

29. The intent manager of claim 28, wherein: The efficiency score comprises a value between 0 and 1, inclusive.

30. The intent manager of claim 29, wherein: The efficiency score comprises a value between L and 1, inclusive, where L is a lower limit of the efficiency score, the lower limit being greater than zero and less than 1.

31. An intent manager according to any one of claims 24 to 30, wherein: The reasoner, based on the efficiency score, invoking the first suggestion agent to generate the suggested action for the environment to solve the problem includes: invoking the first suggestion agent with a probability equal to the efficiency score of the first suggestion agent.

32. The intent manager of claim 31 , wherein: The reasoner invokes the first suggestion agent to generate the suggested action for the environment to solve the problem based on a probability value that is independent of the efficiency score of the first suggestion agent.

33. The intent manager of claim 32, wherein: The reasoner calls the first suggestion agent to generate the suggested action for the environment to solve the problem based on a probability value ε without reference to the efficiency score of the first suggestion agent and a probability value 1-ε based on the efficiency score of the first suggestion agent.

34. An intent manager according to any one of claims 24 to 33, wherein: The efficiency graph indicates efficiency scores of a plurality of suggestion agents, including the first suggestion agent, in generating suggested actions for the environment to solve the problem.

35. The intent manager of claim 34, wherein: The notification indicates whether the suggested actions generated by the plurality of suggestion agents are selected for implementation on the environment, wherein the reasoner adjusts the efficiency score of each of the plurality of suggestion agents based on whether the suggested actions generated by the plurality of suggestion agents are selected for implementation on the environment.

36. An intent manager according to any one of claims 24 to 35, wherein: The first suggestion agent generates a plurality of suggested actions to solve the problem, and wherein the efficiency score of the first suggestion agent is increased in response to any one of the plurality of suggested actions being selected for implementation on the environment.

37. An intent manager according to any one of claims 24 to 36, wherein: The environment includes a computerized system, the information about the environment includes information related to operation of the computerized system, the intent includes a key performance indicator (KPI) of the computerized system, and the problem includes a failure of the computerized system to achieve the KPI.

38. The intent manager of claim 37, wherein: The environment comprises a communication network, the information about the environment comprises information related to operation of the communication network, the intent comprises a key performance indicator (KPI) of the communication network, and the problem comprises a failure of the communication network to meet the KPI.

39. An intent manager according to any one of claims 24 to 38, wherein: The problems include a first problem, and wherein the efficiency map is associated with a situation including a plurality of problems including the first problem.

40. An intent manager according to any one of claims 24 to 39, wherein: The reasoner detects a change in the intent or a change in the question and updates the efficiency score in response to the change in the intent or the change in the question.

41. The intent manager of claim 40, wherein: The reasoner updates the efficiency score by increasing the efficiency score.

42. The intent manager of claim 41 , wherein: The reasoner updates the efficiency score by increasing the efficiency score to a maximum value.

43. The intent manager of claim 41 , wherein: The reasoner updates the efficiency score by increasing the efficiency score by a fixed percentage.

44. The intent manager of claim 40, wherein: The reasoner updates the efficiency score by decreasing the efficiency score.

45. An intent manager according to any one of claims 24 to 44, wherein: The reasoner updates the efficiency score after a predetermined period of time has passed.

46. The intent manager of claim 45, wherein: The reasoner updates the efficiency score by increasing the efficiency score after the predetermined time period has elapsed.

47. The intent manager of claim 45, wherein: The reasoner updates the efficiency score by decreasing the efficiency score after the predetermined time period has elapsed.

Images