@InProceedings{PynadathMarsella2005,
author = {Pynadath, David
V. and Marsella, Stacy C.},
title = {PsychSim: Modeling Theory of Mind with
Decision-Theoretic Agents},
booktitle = {Proceedings of the International Joint
Conference on Artificial Intelligence},
crossref = {},
key = {},
pages = {1181--1186},
year = {2005},
editor = {},
volume = {},
number = {},
series = {},
address = {},
month = {},
organization = {},
publisher = {},
note = {},
annote = {}
}
[1]
Computational modeling of human social behaviour is an important area
of research.
Our beliefs about others (theory of mind) play an important
role within the context of human social interactions and therefore it
should be taken into account when modeling human social behaviour.
Most of the previous computer-simulated models of the theory of mind
have
used first-order logic to represent beliefs and goals. However, such
models have some shortcomings.
On the other hand, PsychSim is one model that generates more plausibly
human behaviour.
PsychSim enables a quick setup of social scenarios within which agents
(representing
individuals) can easily interact with each other.
[2]
These PsychSim agents have their own beliefs about other agents and
their
environments.
They also have goals and policies for achieving those goals.
The PsychSim agents are embedded within a decision-theoretic
framework.
Decision-theoretic frameworks are based on assumptions of rationality
that
people constantly violate. [1]
This framework is an extension to the Com-MTDP model of agent
teamwork.[2]
-MODEL OF THE WORLD:
Each agent starts with a representation of its current state and the
Markovial process by which that state evolves over time in response
to the actions that are performed to change the world:
-State: Includes hidden or obvious objective facts about the
the agent's
world.
Within PsychSim, a vector is used to represent a
state
-Action: Includes an action type, an agent performing the
action , and
another agent that's the object of the action. An
agent
has the option of choosing one action from the
actions
available.
-World: Each action that is performed by the agent changes
the state of
the world.
Probability functions are used to represent the
dynamics of
the world state.
-PREFERENCES:
An agent also has motivation for behaviour as a reward function within
the
decision-theoretic framework. This reward function has 2 components:
1. Minimize/maximize "feature(agent)" goal corresponding to a
negative/positive reward proportional to the value of the
give
state feature.
2. Minimize/maximize "action(actor, object) goal
corresponding to
a negative/positive reward proportional to the number of
actions
performed.
* All these preferences and the relative priority among them
are
represented as a vector of weights.
-BILIEFS:
These computer-simulated agents only have a subjective view about the
world
around them. For instance, the agent A's view of agent B has the same
basic
structure as the real agent B. [2]
However, agent A's subjective view of agent B is considered separate
from the real agent B within the framework and this allows the
representation of errors in beliefs. [3]
Hence, these belief models have a recursive structure, but in reality
people
seldom use very deep optimal behaviour.[2]
An agent's beliefs are affected by action.
[3]
-POLICIES OF BEHAVIOUR:
Each agent's policy is a function representing the process by which it
selects and action based on its beliefs.
This policy is modeled in a way that the agent seeks to maximize
expected
reward of its behaviour.
-MENTAL MODELS:
To simplify the agent's reasoning, the agents' mental models are
realized as
simplifies stereotypes of the richer lookahead behaviour models of the
agents themselves.
These simplifies mental models include potentially erroneous
beliefs about the policies of other agents.
Each agent believes that other agents follow much more reactive
policies as
part of their mental model of each other.
The use of these more reactive policies in the mental models has two
benefits.
First, from a human modeling perspective, the agents perform a
shallower reason
that provides a more accurate model of the real-world entities they
represent.
Second, from a computational perspective, the direct action rules are
cheap to
execute and hence the agents gain major efficiency in their reasoning.
-MODELING INFLUENCE AND BELIEF CHANGE:
Agents use messages to influence the beliefs of one another.
Each message has 4 components: source, recipients, subject and
content.
Messages can refer to beliefs, preferences, behavioural policies and
other
aspects of other agents.
[4]
Influence factors:
- consistency: people expect, prefer and are driven to maintain
consistency between beliefs and behaviour.
- Self-Interest: The inferences we draw and how deeply we
analyze information are all biased by self-interst.
- Speaker's Self-Interest: The tendency of the source of the
message to be more critical if it benefits greatly if
the recipient believes the message.
- Trust, Likeability, Affinity: Trusting, liking or having an
affinity for the message sender greatly affects believing
the message.
These factors can all be modeled by rendering each as a quantitative
function of beliefs that allows an agent to compare alternate
candidate
belief states.
[5]
-PSYCHSIM IN OPERATION:
A good social scene to be modeled by PsychSim is bullying among young
kids.
First generic agent models are selected for various groups or
individuals to
be simulated and specialaized.
These models compute outcome expectancies as the expected value of
actions, meaning that the agent considers the immediate effect of an
act of
aggression and the possible consequences including the change in the
beliefs of other
agents.
A bully is given three subgoals for an act of aggression:
1- To change the power dynamic in the class by making himself
stronger
2- To change the power dynamic by weakening his victim
3- To earn the approval of his peers
Two possible mental models are implemented into the bully's
classmates:
1- Encouraging: they laugh
2- Scared: they laugh only if the teacher doesn't punish them
Three possible mental models of the teacher are implemented into the
bully:
1- Normal: teacher punishes the bully
2- Severe: teacher more harshly punishes the bully
3- Weak: teach doesn't punish the bully
*The relative priorities of these subgoal within the bully's
overall
reward function provide a large space of possible behaviour.
*When creating a model of a specific bully, PsychSim uses a
fitting
algorithm to automatically determine the appropriate weights for
theses goals to
match the observed behaviour
Three specific bully models from the overall space are selected:
1- dominance-seeking
2- sadistic
3- attention-seeking
RESULTS: PsychSim allows one to explore miltiple tactics for dealing
with a
social issue and see the potential consequences.
[6]
-CONCLUSION:
PsychSim is an environment for multi-agent simulation of human social
interaction that employs a formal decision-theoretic approach using
recursive models.
PsychSim has a range of technology within itself that eases the task
of
setting up models.
PsychSim has a range of innovative applications, including
computational
social science and the model of social training environments.