Can AI agents reflect real human emotions and the subtle sense of distance in interpersonal relationships?
Recently, researchers from NVIDIA, the University of Washington, and the University of Hong Kong released Humanoid Agents.
As can be seen from the name, this kind of intelligent agent can reflect the basic needs of human beings.
In the past, intelligent agent simulations did not completely imitate human behavior. The reason is that they did not truly reflect human beings’ basic needs, real emotions, and the subtle sense of distance between people.
The original intention of Humanoid Agents is to develop an intelligent agent that integrates the above elements and is closer to humans.
Agents are guided by System 1 thinking, which responds to specific conditions (such as basic needs), and System 2 thinking, which involves explicit planning.
Currently, the paper has been accepted by EMNLP System Demonstrations 2023.
Paper address: https://arxiv.org/abs/2310.05418
Humanoid Agents
Just as computational simulations of atoms, molecules, and cells have shaped the way we do science, realistic simulations of human-like agents have become a valuable tool for studying human behavior.
You must know that previous intelligent agents had a shortcoming. Although they could complete seemingly believable actions, they were not like real human thinking.
The vast majority of humans do not make plans in advance and then execute these plans meticulously in their daily lives.
In order to alleviate the impact of this shortcoming, researchers drew inspiration from psychology and proposed Humanoid Agents.
Psychologist Kahneman believes that humans have two complementary thinking processes: System 1 (intuitive, relaxed, immediate) and System 2 (logical, intentional, slow).
The Humanoid Agents proposed by the researchers this time introduced the three elements required by System 1 – basic needs (satisfaction, health and energy), emotion and relationship intimacy to make the agent behave more like a human. .
Using these elements, an agent can adjust its daily activities and conversations with other agents.
Moreover, intelligent agents will also comply with Maslow’s needs theory just like humans.
If they don’t interact enough with others, they get lonely; if they don’t stay healthy, they get sick; if they don’t get enough rest, they get tired.
If we only rely on the planning of System 2, we can allow the agent to plan rest time and meet basic needs.
However, without feedback from System 1, the agent cannot take a nap at 3 pm even if it feels tired, because bedtime is scheduled for midnight.
And if the agent feels angry, it needs to do something to vent its emotions, such as running or meditating.
Moreover, the closeness of the relationship between agents will also affect the way they interact with each other.
The social brain hypothesis proposes that our cognitive abilities evolved largely to track the quality of social relationships.
This means that people often adjust their interactions with others based on how comfortable and close they feel to them.
To better imitate humans, the researchers gave the agents the ability to adjust their conversations based on the distance between them.
They proposed a platform that simulates the behavior of humanoid agents in The Big Bang Theory, Friends, and Lin Family, and then visualizes them using the Unity WebGL game interface and uses interactive analytics dashboards to show the agent’s state over time. .
Experiments have proven that Humanoid Agents can effectively respond to and infer changes in various aspects of System 1.
Moreover, this system can be expanded to more aspects, such as personality, moral values, empathy, helpfulness, cultural background, etc.
How it works
In Humanoid Agents, researchers used OpenAI’s ChatGPT-3.5.
Step 1: Initialize the Agent based on the seed information provided by the user.
To put it simply, it is the character setting of each Agent, their name, age, schedule, preferences, etc., and character planning for them.
For example, “John Lin is a Willow Market drugstore owner who likes to help others.” He is characterized by friendliness and kindness.
In addition, the default emotions of Humanoid Agents are set to 7 possible emotions: anger, sadness, fear, surprise, happiness, neutral and disgust.
Step 2: The Agent starts planning his day.
Step 3: Agent takes action according to its own plan. If they are in the same location, Agents can talk to each other, thereby affecting the relationship between them.
The Agent daily plan can recursively decompose the plan at 1-hour intervals, and then at 15-minute intervals to improve the logical consistency of activities over time.
Every 15 minutes, agents perform an activity in their schedule.
However, the Agent can update the plan or make supplements based on changes in internal states, namely emotions and basic needs.
For example, if the Agent is currently very hungry, but the plan is to have a meal in 3 hours.
Here, the Agent can eat some snacks while continuing the current activity. This is especially like the behavior that a worker may have before a meal.
Then, these agents use emotions and basic need states, as well as close relationships with other agents, convert them into natural language descriptions, and then decide how to get along with the agent.
At the end of the conversation, each Agent will look through the conversation history to determine whether they enjoyed the conversation.
If so, their closeness to other agents will be doubled, otherwise, their closeness will be doubled.
Gradual changes in intimacy can be mapped onto how relationships between people develop over time.
In addition, the conversation history is also used to determine whether the agent’s emotions were affected by the conversation.
Step 4: The agent evaluates whether the actions taken have changed their basic needs and emotions.
Step 5: Based on the satisfaction of basic needs and emotions, the Agent can update future plans.
In addition to the default five needs (satiety, socialization, health, entertainment and energy), more basic needs can be added/removed for the Agent.
To do this, users need to create their own default_agent_config.json file in the following format:
{<!-- --></code><code> "name": "fullness", </code><code> "start_value": 5, </code><code> "unsatisfied_adjective": "hungry", </code><code> "action": "eating food", </code><code> "decline_likelihood_per_time_step": 0.05, </code><code> "help": "from 0 to 10, 0 is most hungry; increases or decreases by 1 at each time step based on activity"</code><code>}
The impact of basic needs on activities
Humanoid Agents are dynamic systems made up of many components, so it is a challenge to isolate the impact of each basic need on the agent’s activities.
To investigate the contribution of each basic need, the researchers simulated a world of agents.
These agents initially have a basic requirement set to 0, making the agent extremely hungry, lonely, tired, uncomfortable, or bored at the beginning of the day.
The authors studied the time an agent spent performing activities to satisfy basic needs during a simulated day. For example, eating to overcome hunger and engaging in social activities to relieve loneliness.
The researchers then compared this with the time it would take the Agent to perform such activities under normal conditions (each basic need is set to 5 and energy is set to 10), thereby calculating the time required to satisfy the requirements under test conditions. Percent increase in time spent on each basic need.
As shown in the table below, Humanoid Agents are most adaptable to their activities when basic needs such as health (156%), energy (56%), and satiety (35%) are initialized to 0.
Maslow classified them as low-level “physiological and safety needs” that people need to satisfy before other needs, demonstrating their importance.
In this case, the agent usually sees a doctor, rests, looks for food, etc.
On the other hand, when agents feel lonely due to lack of social interaction, they only slightly adjust their behavior (+12%) to communicate more with other agents.
In addition, another reason why Agent’s social and entertainment activities change less is that under normal circumstances, agents have spent a lot of time doing activities that satisfy these basic needs.
On average, they spend 11 hours doing things they enjoy, 8.75 hours on social interactions, only 5.75 hours on resting, and 2.75 hours on eating and doing things to improve their health.
This means that the effect of initially setting fun or socializing to 0 disappears very early in the day and is replaced by other priorities, including work obligations, such as Penny working at the cheesecake shop.
The impact of emotions on activities
Here, the authors studied the number of times (15-minute intervals) that the agent performed activities expressing each emotion during a single day of simulation.
For example, when the Agent is angry, he will run to vent his anger; when he is sad, he will seek a trustworthy friend to talk to; when he is disgusted, he will practice deep breathing and meditation; when he is surprised, he will take time to Process and ponder these surprising findings.
The researchers then calculated the difference in the number of times the agent performed such activities in a normal setting.
Under normal circumstances, Agents usually do not show behaviors of sadness, anger, fear, disgust, or surprise, but experimental results show that the number of behaviors expressing these emotions has increased compared with Agents under normal circumstances.
As shown in the table above, anger has the greatest impact on agents (+15 activities), followed by sadness and fear (+10 each), then disgust (+4) and surprise (+1), and finally happiness (-2) .
Negative emotions appear to affect agents more than positive emotions because agents typically do not plan to do activities with negative emotions and therefore have to significantly adjust their plans to manage negative emotions.
Interestingly, it was observed that when the Agent was happy, he did less activities in order to keep himself happy.
The impact of intimacy on activities
In addition, the author also studied the impact of initial relationship intimacy on the dialogue between two agents.
As shown in the table below, as intimacy increases, the average number of turning points in a conversation takes an inverted U shape.
The Agent speaks less when the distance is far away, and speaks more when the distance is close, but when the distance is very close, it will gradually decrease again.
This is very similar to humans, when we feel very close to others, there is less need for polite conversation.
Interestingly, in Lin’s Family, this turning point occurs at “rather close”. In Friends and The Big Bang Theory, the turning point occurs at “close”.
Perhaps it is because the two agents in Lin have a father-son relationship. It is comfortable to have less communication at a lower level of intimacy and will not strain the relationship.
In Friends and The Big Bang Theory, agents are friends and neighbors with each other and require more active communication to maintain relationships.
Comparison with human annotation
In order to evaluate the predictive ability of Humanoid Agents (such as whether the activity satisfies one’s basic needs, what emotions one will show during the activity, and whether the conversation will bring two agents closer), the researchers compared the system’s predictions with human annotations.
Three human taggers will use the same commands as ChatGPT to tag a simulation of a day in the Lin family world.
Each annotator independently annotated 144 activities for emotional and basic needs, and 30 annotations for user dialogue pairs.
We then conducted a majority vote for all annotators and calculated the micro-F1 between the majority vote and the system prediction.
Table 1 shows that inter-rater reliability was good (Fleiss’ κ >= 0.556) for all basic needs, emotions, and relationship intimacy.
The researchers also found that if an activity increases satiety and energy, the Agent can perform well in classification (F1 >= 0.84).
Moreover, agents can express emotions during activities, and conversations can also shorten the distance between different agents.
However, the Agent performed poorly when assessing whether the activity met the basic needs of fun, health, and social interaction.
The reason may be that the agent system greatly overestimates the number of activities required to meet these needs. For example, health accounts for 34% of Agent predicted activities vs. 4.9% of human labeled activities, entertainment accounts for 44.4 vs. 10.4%, and social networking accounts for 47.2% vs. 24.3%.
The Agent will think that because John Lin works in a pharmacy, these activities will help the Agent’s health; receiving feedback from the professor or helping old customers buy medicines will make him happy.
The researchers judge that this problem may be alleviated if a language model that better understands common sense is used.
AI agents are all built using LLM, where is the innovation?
As soon as Stanford’s landmark paper on Westworld Town came out, a lot of imagination was stimulated in the industry, and research on using LLM to build imaginable human behavior intelligence agents also emerged one after another.
What is the difference between Humanoid Agents and previously popular agents such as BabyAGI and AutoGPT?
Researchers pointed out that Humanoid Agents should be the only job currently simulating the daily activities of humanoid agents. Other jobs generally achieve externally defined goals.
For example, Langchain Agents, BabyAGI, AutoGPT, AgentVerse, Voyager, and CAMEL all build task-oriented agents and solve user-defined tasks by recursively decomposing them into simpler subtasks.
Moreover, the multiple rounds of dialogue responses generated based on emotions, game character descriptions, and personal facts are not dynamically simulated by the agent, but are based on a set of static, character-independent text information.
These previous works cannot simulate the impact of dynamic attributes, such as changes in intimacy between a pair of agents.
Moreover, Humanoid Agents can take into account basic needs, emotions, intimacy and other aspects at the same time when generating conversational responses, just like real humans.
In previous work, only one aspect was considered at a time.
Customized Agent
Currently, the system supports three built-in settings:
1. Big Bang Theory
--map_filename ../locations/big_bang_map.yaml \ --agent_filenames ../specific_agents/sheldon_cooper.json ../specific_agents/leonard_hofstadter.json ../specific_agents/penny.json
2. Friends
--map_filename ../locations/friends_map.yaml \ --agent_filenames ../specific_agents/joey_tribbiani.json ../specific_agents/monica_gellor.json ../specific_agents/rachel_greene.json
3. Lin Family
--map_filename ../locations/lin_family_map.yaml \ --agent_filenames ../specific_agents/eddy_lin.json ../specific_agents/john_lin.json
At the same time, users can also create their own maps and agents through customized settings.
It should be noted that the agent and the map are not completely separated. For each agent_filename specified, its name field must be included as a key in map.yaml under Agents.
Analytics Panel
The data generated by the Agent during its activities can be visually displayed through an interactive dashboard. It includes a basic needs map and a social relationship map, along with corresponding information including emotional and conversational details. ?
cd humanoidagents</code><code>python run_dashboard.py --folder <folder/containing/generation/output/from/run_simulation.py>
Required parameters
-
–folder is the folder where the output results generated by run_simulation.py are stored.
-
–mode is a method for selecting data from a folder. It has two modes: (1) all: display all files in the folder (2) date_range: display files within the selected date range (needs to be specified in the parameters)
Optional parameters
-
When –mode = date_range, –start_date is the starting date (inclusive). The format is YYYY-MM-DD, such as 2023-01-03
-
When –mode = date_range, –end_date is the end date (inclusive). The format is YYYY-MM-DD, such as 2023-01-04
Author introduction
Zhilin Wang
Zhilin Wang is a senior application scientist on NVIDIA’s NeMo NLP team. Previously, he received a master’s degree from the University of Washington, studying natural language processing, researching dialogue systems and computational social science.
References:
https://github.com/HumanoidAgents/HumanoidAgents
The knowledge points of the article match the official knowledge files, and you can further learn relevant knowledge. Python entry skill tree Home page Overview 381,219 people are learning the system