Information processing system

Abstract

To improve the simulation of human behavior in terms of accuracy and / or computational efficiency. [Solution] The information processing system stores instructions for generating an agent, and the agent has an agent memory that stores historical information related to the agent, a first large-scale language model that uses persona data characterizing the agent's persona as a condition, a second large-scale language model, and a control module, and the control module performs the following: causes the first large-scale language model to generate one or more actions based on the historical information, causes the second large-scale language model to output an evaluation of the one or more actions generated by the first large-scale language model, and causes the first large-scale language model to update one or more actions based on the evaluation output by the second large-scale language model.

Description

Technical Field

[0001] The present disclosure relates to an information processing system, and more particularly to an information processing system for simulating human behavior.

Background Art

[0002] Techniques for predicting human behavior based on historical information are known. For example, Patent Document 1 discloses a behavior prediction device that predicts human behavior using a prediction model learned based on human behavior history information.

Prior Art Documents

Patent Documents

[0003]

Patent Document 1

Summary of the Invention

Problems to be Solved by the Invention

[0004] Since the simulation and prediction of human behavior are useful, techniques for improving the simulation of human behavior are desired.

[0005] The information processing system according to this disclosure comprises a control unit and a storage unit, the storage unit storing instructions that cause the information processing system to generate an agent when executed by the control unit, the agent comprising an agent memory that stores historical information related to the agent, a first large-scale language model that is conditional on persona data characterizing the agent's persona, a second large-scale language model, and a control module, the control module performing the following actions: causing the first large-scale language model to generate one or more actions based on the historical information, causing the second large-scale language model to output an evaluation of the one or more actions generated by the first large-scale language model, and causing the first large-scale language model to update one or more actions based on the evaluation output by the second large-scale language model. [Effects of the Invention]

[0006] The information processing system disclosed herein enables improved accuracy in simulating human behavior. [Brief explanation of the drawing]

[0007] [Figure 1] This is a schematic diagram showing an example of the configuration of an information processing system. [Figure 2] This is a schematic diagram showing an example of an agent. [Figure 3] This is a flowchart illustrating an example of a process for simulating human behavior. [Modes for carrying out the invention]

[0008] Figure 1 is a schematic diagram showing an example configuration of an information processing system 10 according to one embodiment. The information processing system 10 includes a control unit 12 and a storage unit 14. The control unit 12 includes one or more processors, and the storage unit 14 includes one or more memory units. The information processing system 10 shown in Figure 1 is simplified for ease of understanding, but may include one or more network interfaces, one or more communication interfaces, one or more input interfaces, and / or various other equivalent components. For example, the storage unit 14 stores instruction information 16, which, when executed by the control unit 12, implements one or more functions of the information processing system 10. The instruction information 16 may be supplied to the information processing system 10 on a non-transient computer-readable medium, or it may be received via a network (not shown) such as the Internet, a mobile communication network, an ad-hoc network, a local area network (LAN), a metropolitan area network (MAN), or any combination thereof.

[0009] The information processing system 10 simulates or predicts human behavior using artificial intelligence. More specifically, the information processing system 10 generates software agents (hereinafter simply referred to as "agents") that simulate one or more human behaviors based on persona data that characterizes a desired persona for the agent. Furthermore, the information processing system 10 may be configured to simulate or predict human behavior based on agent memory that stores historical information characterizing one or more actions previously performed by the agent.

[0010] Figure 2 is a schematic diagram showing the agent 100 instantiated by the information processing system 10 in Figure 1. For ease of explanation, Figure 2 illustrates the agent 100 by representing each individual unit of functionality implemented in software by the agent 100 as a function. The functions shown in Figure 2 can be combined and / or divided as appropriate.

[0011] Agent 100 includes a control module 102, agent memory 104, a first large-scale language model (LLM) 106, a second LLM 108, persona data 110, one or more subordinate control modules 112, and a descriptor module 114.

[0012] Agent 100 interacts with the simulation environment 200. The simulation environment 200 may be implemented by the information processing system 10 or by a separate system outside of the information processing system 10. The simulation environment 200 is a virtual environment in which Agent 100 interacts to perform various actions or tasks. For example, the simulation environment 200 may be a simulation of a town or city in which Agent 100 resides. In some embodiments, multiple agents interact with the simulation environment 200, thereby enabling multiple agents to interact with each other. Interaction with the simulation environment 200 can be achieved through an application programming interface (API) provided by the simulation environment 200.

[0013] The control module 102 is configured to provide overall control of agent 100. Specifically, the control module 102 is configured to control or organize operations performed by the first LLM 106, the second LLM 108, one or more subordinate control modules 112, and the descriptor module 114 to generate one or more actions. The control module 102 can perform control by instructing each of the first LLM 106, the second LLM 108, one or more subordinate control modules 112, and the descriptor module 114 according to a control procedure, thereby instructing each component of agent 100 to perform a specific task. Such instructions may be provided according to a specific syntax, API, and / or in natural language form. For example, the control module 102 can control the first LLM 106 and the second LLM 108 by issuing one or more natural language prompts.

[0014] Persona data 110 specifies or characterizes the human being simulated by agent 100. That is, persona data 110 encapsulates the human characteristics of agent 100, enabling agent 100 to simulate human behavior in a realistic manner. Persona data 110 may include one or more parameters that characterize agent 100's persona, and / or one or more natural language descriptions of agent 100's persona. For example, persona data 110 may include one or more parameters that characterize the personality, physical attributes, interests, and / or background of the human being simulated by agent 100. Here, personality may be specified by one or more parameters or scores corresponding to one or more personality attributes. Alternatively or additionally, personality may be specified in terms of natural language descriptions.

[0015] Persona data 110 may be generated based on historical information obtained from a group of people. For example, such historical information may be obtained through a system that monitors the activities of a group of people. In another example, the historical information may be obtained from historical information generated by agent 100 itself. Based on this historical information, a summary of personality attributes is generated in natural language. This summary is used by LLM to generate persona candidates that match the historical information. For example, LLM may be instructed to generate multiple candidate personas and provide a score for each candidate persona indicating consistency with the historical information. In some cases, to increase the diversity of persona candidates, LLM may be provided with a set of selectable traits (e.g., a set of different occupations). Once an appropriate persona is selected, the selected persona is stored in persona data 110.

[0016] Agent memory 104 stores historical information of agent 100. This historical information provides a record of agent 100's past actions and experiences. Agent memory 104 is replenished in real time as agent 100 interacts with the simulation environment 200. Alternatively, or additionally, agent memory 104 may be initialized based on historical information generated by agent 100 in previous simulations. Agent memory 104 functions as a long-term record of agent 100's actions, observations, emotions, and thoughts, and thus characterizes agent 100's internal state.

[0017] The first LLM 106 functions as a high-level planner LLM and is configured to generate a daily plan that includes one or more actions to be performed by agent 100. Specifically, the first LLM 106 generates a daily plan that matches the persona associated with agent 100, conditionally based on persona data 110. The first LLM 106 may refer to agent memory 104 to ensure that the daily plan matches agent 100's history. Typically, the first LLM 106 may generate a daily plan on a daily basis (i.e., once a day). The daily plan can specify one or more actions to be performed according to one-hour time slots, along with where each action will be performed and any subtasks associated with each action.

[0018] The first LLM106 outputs a daily plan to the second LLM108. The second LLM108 functions as a critic LLM and is configured to evaluate the daily plan generated by the first LLM106. More specifically, the second LLM108 is configured to determine whether the daily plan generated by the first LLM106 is realistic and matches the persona defined by the persona data 110 and / or the historical information stored in the agent memory 104. The second LLM108 also determines whether the daily plan matches one or more external contexts, such as the simulation environment 200, cultural norms, and / or human habits. The second LLM108 outputs an evaluation of the daily plan to the first LLM106. The evaluation includes a score for the daily plan and / or natural language feedback on the daily plan.

[0019] Upon receiving feedback, the first LLM 106 updates or modifies the daily plan as necessary. By refining the daily plan based on feedback from the second LLM 108, the first LLM 106 generates a realistic daily plan that is conditional on the persona data 110 and matches the internal state of agent 100, characterized by historical information stored in agent memory 104. In this way, agent 100 can generate realistic human behavior.

[0020] The first LLM106 and the second LLM108 may be implemented using a pre-trained LLM, such as Generative Pre-trained Transformer 4 (GPT-4) created by OpenAI, Inc. in San Francisco, California; or by fine-tuning a pre-trained LLM; or by training an LLM specifically for use as the first LLM106 and / or the second LLM108.

[0021] After the daily plan is generated, criticized, and updated as needed, it is executed in the simulation environment 200. The execution of the daily plan is delegated to one or more lower-level control modules 112. One or more lower-level control modules 112 are configured to execute one or more low-level tasks required for the execution of the daily plan. For example, one or more lower-level control modules 112 include a controller specialized in predicting the optimal way to move between two locations within the simulation environment 200. One or more lower-level control modules 112 can include one or more LLMs trained to execute specific low-level tasks and / or one or more models based on action trees, reinforcement learning (RL), neural networks, etc. Usually, one or more lower-level control modules 112 execute low-level tasks based on defined policies without considering the persona of the agent 100. Thus, by delegating the low-level tasks required to execute the daily plan to one or more lower-level control modules 112, the agent 100 can simulate human behavior with improved computational efficiency.

[0022] The execution of the daily plan in the simulation environment 200 can include several steps. First, the agent 100 can make one or more observations of the simulation environment 200. These observations may be made by one or more lower-level control modules 112 that control the interaction between the agent 100 and the simulation environment 200. In some embodiments, these observations may be visual observations translated into natural language by the descriptor module 114. In such embodiments, the descriptor module 114 can be implemented by a vision-language model (VLM) configured to caption the visual observations made by one or more lower-level control modules 112 and provide the captions to the first LLM 106 and / or the second LLM 108 in the form of natural language text.

[0023] When the first LLM 106 receives observation results from the lower control module 112 and / or the descriptor module 114, it determines whether one or more actions to form the daily plan need to be changed. For example, based on the observation results, the first LLM 106 estimates the current mood of the agent 100 and the physical state of the agent 100 (e.g., hunger, fatigue), and determines whether a modification of the daily plan is necessary. For example, if the first LLM 106 determines that the agent 100 is hungry and the daily plan does not involve a meal for several hours, the first LLM 106 decides to go to a restaurant in the simulation environment 200 and accordingly modifies the daily plan. In a further example, if one or more observation results indicate that it is raining at the current location of the agent 100 in the simulation environment 200 and the daily plan includes one or more outdoor activities, the first LLM 106 determines that a modification of the daily plan is necessary.

[0024] One or more observations of the simulation environment 200 can also be used to update the historical information stored in the agent memory 104 so that the historical information accurately reflects the experience of the agent 100 in the simulation environment 200. In this way, the experience of the agent 100 in the simulation environment 200 can be reflected in the generation of future daily plans by the first LLM 106, thereby enabling a more accurate simulation of human behavior.

[0025] In some embodiments, the agent 100 can provide an interface (not shown), such as an API, that enables questions to the first LLM 106 via one or more natural language prompts. For example, a human operator can use the interface to ask the agent 100 why it performed or is performing a particular action in the simulation environment 200. In this way, the human operator can gain insights into the persona of the agent 100. Such insights can be beneficial for urban planning regarding the town or city simulated by the simulation environment 200.

[0026] Figure 3 is a flowchart showing procedure 300 for simulating human activity in one embodiment. For clarity, the following description assumes that procedure 300 is performed by the control module 102. However, procedure 300 may be performed by any component of agent 100, or any combination of components of agent 100. Here, procedure 300 is an example of the control procedure described above with reference to Figure 2.

[0027] In 302, the control module 102 causes the first LLM 106 to generate one or more actions based on the history information stored in the agent memory 104. As described above, one or more actions may constitute a daily plan. Here, since the first LLM 106 is conditioned based on the persona data 110, one or more actions reflect the persona and past experiences of agent 100.

[0028] In 304, the control module 102 causes the second LLM 108 to output an evaluation of one or more actions generated by the first LLM 106. As described above, the evaluation may include a score for one or more actions and / or feedback about one or more actions in natural language format. This evaluation indicates whether one or more actions are realistic and match the persona defined by the persona data 110 and / or the history information stored in the agent memory 104.

[0029] In 306, the control module 102 causes the first LLM 106 to update one or more actions based on the evaluation output by the second LLM 108. This ensures that one or more activities are realistic for the persona defined by the persona data 110 and are consistent with the internal state of agent 100, characterized by the history information stored in agent memory 104.

[0030] In 308, the control module 102 causes one or more actions to be performed in the simulation environment 200. As described above, the execution of the daily plan can be achieved by one or more sub-control modules 112.

[0031] In 310, the control module 102 causes the first LLM 106 to update one or more actions based on one or more observations of the simulation environment 200. As described above, these observations may be made by one or more sub-control modules 112 that control the interaction between agent 100 and the simulation environment 200. In this way, agent 100's experience in the simulation environment 200 can be used to improve one or more actions generated by the first LLM 106.

[0032] Embodiments of the present disclosure have been described with reference to the drawings and examples, and various modifications and revisions can be made by those skilled in the art based on this description. Accordingly, such modifications and revisions are included within the scope of the present disclosure.

Claims

1. An information processing system having a control unit and a storage unit, The storage unit stores instructions that cause the information processing system to generate an agent when executed by the control unit. The agent in question, Agent memory that stores historical information related to the agent, A first large-scale language model that is conditional on persona data characterizing the persona of the agent, The second large-scale language model, It has a control module, The control module is The large-scale language model described in 1 above generates one or more actions based on the historical information, The second large-scale language model is made to output evaluations of one or more actions generated by the first large-scale language model, The first large-scale language model is made to update one or more actions based on the evaluations output by the second large-scale language model. Execute Information processing system.

2. In claim 1, The control module is A step of performing one or more actions in a simulation environment, The first large-scale language model is instructed to update historical information based on one or more observation results obtained from the simulation environment. Information processing system.

3. In claim 1, The agent has a descriptor module configured to convert one or more observation results obtained from the simulation environment into natural language descriptions of one or more observation results, The control module causes the first large-scale language model to update one or more actions of the first large-scale language model based on the natural language descriptions of one or more observations. Information processing system.

4. In claim 2, The agent has one or more subordinate control modules configured to perform one or more actions in the simulation environment. Information processing system.

Images