A Bayesian Game Theory Decision Model of

Resource Optimization for Emergency Response

Mark Rahmes, Kevin Fox, Kevin Davis, Brian Griffin

Harris Corporation,
Government Communications Systems Division
Melbourne, Florida 32904

Abstract─ We describe a system model for determining decision Economic factors (e.g. unemployment rates, prices for
making strategies based upon the ability to perform data mining food, such as bread, or fuel), Political factors (freedoms, type
and pattern discovery utilizing open source information to of government), Religious factors (type of religions, religious
prepare for specific events or situations from multiple
information sources. Within this paper, we discuss the
tensions) combined with trend information such as sentiment
development of a method for determining actionable information. analysis on social media, open source data, news, etc. can
We have integrated open source information linking to human provide indicators of areas undergoing stress or at risk.
sentiment and manipulated other user selectable interlinking Current situational awareness requires efforts to seek to
relative probabilities for events based upon current knowledge. incorporate not only geospatial features and forces structures,
Probabilistic predictions are critical in practice on many decision but also the human element, especially in urban settings. An
making applications because optimizing the user experience attempt to predict the likelihood of reaction to a future event
requires being able to compute the expected utilities of mutually
exclusive pieces of content. Hierarchy game theory for decision
will be based on correct situation analysis. Efforts to combine
making is valuable where two or more agents seek their own the information required for these predictions are time
goals, possibilities of conflicts, competition and cooperation. The consuming and labor intensive. The availability of open source
quality of the knowledge extracted from the information available social media information and implementation of artificial
is restricted by complexity of the model. Hierarchy game theory intelligence (AI) methodologies makes this problem tractable.
framework enables complex modeling of data in probabilistic Our GlobalSite system, shown in Fig 1, can also be used as a
modeling. However, applicability to big data is complicated by the method for asset management and reduce cost of analyses.
difficulties of inference in complex probabilistic models, and by
computational constraints. We focus on applying probabilistic
models to resource distribution for emergency response.
Hierarchical game theory models interactions where a situation
affects players at multiple levels. Our paper discusses the effect of
optimizing the selection of specific areas to help first responders
and determine optimal supply route planning. Additionally we
discuss two levels of hierarchies for decision making including
entry decisions and quantitative Bayes modeling based on
incomplete information.
Index Terms—Game theory, Resource Management, Decision
Making, Operations Research

I. INTRODUCTION
Game theory is the study of strategic decision making. It is
the study of mathematical models of conflict and cooperation
between intelligent rational decision-makers and is often
thought of as an interactive decision theory. It has been applied
to economics, political science, psychology, logic, biology and
other complex issues. Modern game theory began with the idea Fig1. Overview
regarding the existence of mixed-strategy equilibrium in two-
person zero-sum games, applied to economics. Later this As an example, consider the recent case of Typhoon
evolved to provide a theory of expected utility, which allowed Haiyan, which devastated portions of the Philippines in early
mathematicians and economists to treat decision-making with November 2013. Weather data and hurricane/typhoon forecast
uncertainty. The notion of probabilistic predictions utilizing models could be used to project the path of the storm, and
game theory is critical in practice to many decision making anticipate areas that may be affected. This could lead to
applications because optimizing user experience requires being enriching Foundation GEOINT content for the Philippines in
able to compute the expected utilities of mutually exclusive anticipation of the event (landfall of Typhoon Haiyan), as well
pieces of data which is critical to geospatial analytics. as collection of additional data after the event to detect
changes, assess damage, and support Disaster quicker than if relief workers just ran into the Philippines with
Relief/Humanitarian Aid. For instance, change detection may no preparation or information [9].
reveal roads have been washed out, presenting logistics The crowdsourcing involved people from all around the
problems for the delivery of aid to folks in need. world who viewed satellite images from space and provided
The Philippines has strategic importance to the U.S. as part relief agencies with their knowledge of the changes that had
of the strategy plans to counterbalance China’s rising military occurred on the ground after the storm passed. Officials from
influence with strong American allies in the region. The U.S. the United Nations Office for the Coordination of
and the Philippines are in the middle of negotiating an Humanitarian Affairs (OCHA) coordinated the effort to get
increased American military presence in the country [8]. volunteers to help with the aid relief. Doctors Without Borders
received updated maps generated by over 1,000
II. OPEN SOURCE DATA OpenStreetMap volunteers in 82 countries. They identified
The internet has forever changed the way people are able hospital locations, which buildings were intact and which were
to respond to a disaster. Now, a person, business, damaged, blocked roads, and other key infrastructure [10].
or organization can create a call to action that generates Technological advances in sensing, computation, storage,
millions of dollars’ worth of donations in money, food, and and communications will turn the near-ubiquitous mobile
even volunteer power in a matter of minutes. This can happen phone into a global mobile sensing device. People-centric
via an email, a button on a website, or a YouTube video that sensing will help drive this trend by enabling a different way to
goes viral. We have seen this during disastrous events like sense, learn, visualize, and share information about ourselves,
Hurricane Katrina, the 2010 earthquake in Haiti, or the recent friends, communities, the way we live, and the world we live
typhoon in the Philippines. The word, “crowdsourcing,” is a in. It juxtaposes the traditional view of mesh sensor networks
combination of two words, crowd and outsourcing. Thus with one in which people, carrying mobile devices, enable
crowdsourcing, as it applies to disaster response, is the process opportunistic sensing coverage [3].
Since people centric sensing began, content provided by
of gathering work or funding via the internet to benefit a
ordinary people, so-called "citizen journalists" or individuals
particular person, organization, or event [9].
with particular agendas that is posted or shared on Social
What makes crowdsourcing so important is the belief that Networks such as Twitter, YouTube, Facebook, MySpace or
more heads are better than one. Using the canned food drive as Flickr, to name but a few, has increasingly made it into the
an example, if you were to do the work without the internet, channels and services of traditional information providers such
you would have to run around town to various homes and as news organizations. New and affordable publishing and
businesses and ask individuals if they would like to participate. distribution tools for ordinary citizens such as Social
This would take up too much time and man power. The Networks, blogs, or services have made this possible. Social
internet can be used to send email to friends, who would then Networks have more and more become an integral part of the
pass the word on to their friends. An online donation campaign communication mix for all kinds of aims, for example
can be created where one can make a short video as to why (political) campaigning, and awareness-raising [4]. See for
people should donate to a cause [9]. example, Fox News revamped its newsroom for Shephard
The recent typhoon in the Philippines has seen an exciting Smith Reporting on breaking news, such as December 2013
change in how crowdsourcing can assist in disaster response. shooting at Arapahoe High School in Colorado. Open source
Rather than sit and wait for heads of organizations and data is valuable in order to populate the reward matrices for
governments to dictate what is needed on the ground, people game theory applications.
are able to assist first responders in the very work of saving
lives, both directly and indirectly. Through the use of powerful III. GAME THEORY
technology, people are able to track weather patterns that are Current situational awareness efforts seek to incorporate
more accurate than anything you will find on the evening not only geospatial features and structures, but also the human
news. Geography buffs are able to use satellite imaging element, especially in urban settings. An attempt to predict the
technology to create maps and locate where people are likelihood of human reactions to a future event should be based
stranded and in desperate need of food and water. There are on correct situational analysis. Development of tools for more
even examples of people who have been able to locate others rapid refinement of flexible plans is required for adapting to a
who were buried under debris. This kind of response is a much changing operational environment.
more aggressive response to a disaster [9]. Our solution populates a reward matrix in near real time
Social media tools like Twitter and Facebook, traditionally through powerful game theory analysis. Once data accuracy is
looked upon as a game for kids has been useful to relief proven through sensitivity analysis, the information is can
workers as well. The group Standby Task Force has been able either be used as training data or populated into a reward
to gather over a million tweets, text messages, and other social matrix in real time for resource allocation and adversarial
media updates to track the extent of the damage in near real planning utilizing game theory analysis. Our techniques enable
time. They were able to create a map using the assistance of a methodical approach to intelligent planning and reaction
hashtags that allowed them to gather the information much based upon construction and analysis of a decision model
resulting in a structure of the most probable solution. This IV. EMERGENCY RESPONSE EXAMPLE
technique is useful for a number of applications ranging from
A. Resource Planning
behavioral economics, war fighter planning, and analysis of
information, messaging, and risk management. Our system In our example, there are several resource management stages
supports an artificial intelligence (AI) supervised learning or hierarchies as shown in Fig 2. These stages include information
needs, collection objectives, observables, tasks and plans. The
approach to quantify information based on user selectable
resource management process seeks to decompose information
attributes and deriving probabilistic decision outcomes. Our
needed to satisfy mission objectives into one or more tasks. The
approach trains with near real time execution.
essence of resource management is uncertainty management [13].
Our solution integrates multiple data sources into efficient Resource allocation problems in which limited resources must
intent analysis processes and uses training data to build the be allocated among several activities are often solved by
decision trees to predict categories for new events based upon dynamic or linear programming. Operations Research is a
classifiers created for the use case scenario. Given an event, we branch of mathematics that studies decision making to obtain
predict a category and then determine sentiment based on the best decision. Game theory can help determine the optimal
trained data. This information could then be applied during investment strategy [24].
planning in support of course of action (COA) development in
the military decision making process (MDMP).
The approach combines the following input: open
(unstructured) source, and/or direct user input/modification. In
particular, we capture and model “sentiment” and other
situational factors through the assignment of positive, neutral
and negative values. A reward matrix is then populated using
game theoretic concepts such as in a competitive game model.
GlobalSite utilizes game theory which permits the ability to
solve for iterative solutions, instantaneous visual feedback, and
interactions by the user on demand. Our output enables a
methodical approach to intelligent planning and reaction
including interaction of variables, parameters and attributes by
user resulting in updated probabilities. Game theory is useful
for resource management of manpower, equipment, and
warnings, etc., since it shows optimal decision for deployment. Fig 2. Bayesian Hierarchy
In many situations, the opponents know the strategy that
they are following. We assume that the players know what Our solution populates a reward matrix in near real time
actions are available. A maximin equilibrium often is the through powerful game theory analysis. Once data accuracy is
strategy and is called the Nash theory application of zero or proven through sensitivity analysis, the information can either
constant sum strategy game. We also consider a constant sum be used as training data or populated into a reward matrix for
game in which for both player’s strategies, the two player’s resource allocation and adversarial planning utilizing game
reward add up to a constant value. This means, while both theory concepts such as in a competitive or cooperative game
players are in conflict, that there is more to gain than simply model. Much of the current focus is on human geography and
having one player’s reward equaling the other player’s loss. terrain as well as population based sentiment analysis [17].
We can find optimal strategies for this two-person zero-
sum game [24]. For example, if a reward matrix exists, then
the equilibrium point is the one where the reward is the
smallest value in its row and the largest number in its column.
A pure strategy provides a complete definition of how a player
will play a game. A player's strategy set is the set of pure
strategies available to that player. A mixed strategy is an
assignment of a probability to each pure strategy. This
equilibrium is also known as the Nash Equilibrium [15].
Game theory is divided into two branches, non-
cooperative and cooperative [2]. Algorithms for computing
Nash equilibrium are well-studied. N-player games are
computationally harder than 2-player games, in important ways Fig 3. Reward Matrix
such as visualization of the solution [11].
Figure 3 shows populated example values for a resource
planning game. We use the Nash equilibrium to solve for the
mixed solutions in a repeatable and methodical manner to weights to a file allows for peer review in order to check and
determine optimal choices. In our example, open source data is validate decisions. Our approach is modeled, so that the
used to create a cost function. In our example using the reward process can be repeated to allow for new or higher quality
matrix, we show the linear programming solution for the data/information to be inserted into the process to generate
constant sum game as follows: updated results.

max v (1) B. Path Planning

s.t. v - 372 x1 - 16 x2 - 7 x3 - 2 x4 <= 0 Path planning algorithms are commonly used to find the
v - 181 x1 - 2 x2 - 1 x3 - 1 x4 <= 0 least cost path from a start node to an end node through a
v - 160 x1 - 2 x2 - 1 x3 - 0 x4 <= 0 gridded environment of cost. The cost of a given node-to-node
v - 100 x1 - 50 x2 - 20 x3 - 1 x4 <= 0 transit may simply relate to the distance traveled combined
v- 40 x1 - 20 x2 - 5 x3 - 2 x4 <= 0 with some measure of an obstruction. This allows the
v- 12 x1 - 4 x2 - 2 x3 - 1 x4 <= 0 algorithm to find the shortest path through a maze for example.
x1 + x2 + x3 + x4 = 1 Given a weighted directional graph with a start node and a set
x1, x2, x3, x4 >= 0 of end nodes, the optimal path problem is to find a least-cost
path from start to any member of end, where the cost of the
Equation (1) is used to determine the best strategy for the path may, in general, be an arbitrary function of weights
blue player to deliver aid. The solution for the blue player’s assigned to nodes and branches along path.
mixed strategy in terms of probabilities x = (1, 0, 0, 0). Figure 4 Our algorithm applies an additive evaluation function:
shows iterative modeling of the situation over time periods in
order to optimize decision making. Initially, the best decision is 𝑓(𝑛) = 𝑐𝑡(𝑛) + s(𝑛) + 𝑒(𝑛) + h(n) + 𝑑(𝑛) (2)
to send aid to Eastern Visayas in accordance with the reward
where 𝑐𝑡(𝑛) is the cost of terrain movement due to damage to
matrix.
infrastructure evaluated along dynamic path in north direction.
The safety value 𝑠(𝑛) is the cost of the safety along dynamic
path in north direction. Elevation information value 𝑒(𝑛)
derived from SRTM or LiDAR data is the cost in the dynamic
path in the north direction. The hospital presence value ℎ(𝑛) is
the cost in the dynamic path in the north direction. The
distance value 𝑑(𝑛) is the minimum possible cost to reach the
end from a given node and is a crucial component of the
algorithm. It drives the algorithm towards the end node. It
should be noted that𝑓(𝑒),𝑓(𝑤) and 𝑓(𝑠) correspond to nodes
in the east, west, and south directions respectively.
In order to avoid a combinatorial explosion of potential
Fig 4. Iterative Modeling paths, the algorithm, and therefore target motion, is restricted
to a gridded environment of nodes. This also ensures that the
When the reward matrix contains no saddle point, we can use
same node may be revisited from a different direction allowing
a linear program solver. Some tools use “strategies” measured in
for the best path to be continuously updated. The distance
different units in the same reward matrix and can be
between nodes should be as large as possible in order to
problematic. If all strategies in a given decision model reward
minimize the number of nodes that must be expanded, thereby
matrix are not in the same (equalized) units, then use of game
reducing computation time, while not impacting on the
theory and mini-max or maxi-min functions can provide
solution [23].
misleading results. We can create purely dominant and
incorrect solutions just due to relative size of unit measures.
Our solution addresses this properly and uniformly for any
decision model. We equalize all strategies (in a given decision
model) to the same unit. This is a key point to the application
of game strategies to a general class of decision problems. An
adjustable “equalization” factor has the purpose to convert all
strategy measures to the same unit (e.g., cost, time) and must
be done for any decision model. The equalization factor for our
solution is independent of additional (importance) weights that Fig 5. Eight Connected Movement Direction Methods
may be applied.
Using different weights allowed for choices is to highlight The resolution of the grid we chose is 90 meters on a side.
the ability and need for a tool which can be used to allow the Figure 5 shows two possible movement methods through the
user to dial and modify modeled parameters of the reward grid. The eight connected method lets the user travel
matrix to model “what if” scenarios. Additionally saving the diagonally. This does mean extra compute time per movement,
but yields more realistic results. The goal is to minimize cost
of traveling along an optimal path. We are seeking a minimum
cost function from the reward matrix shown in Table 1, with
eight connected movements.
Table 1. Reward Matrix

Fig 6. Optimal Path to Drop Off Center

More than one million Tweets and text messages were

tapped and then mapped using MicroMappers, applications
designed for disaster response. In Tacloban, the city that bore a
direct assault from Haiyan’s storm surge, 58% of the built up
areas were destroyed or damaged. In the image below you can
see orange, red and yellow color-keyed buildings. The orange,
shown in Fig 7b, represent seriously damaged structures. The
red indicates buildings that were completely destroyed. The
It is acceptable to model a player’s choice of strategies with yellow are buildings largely intact [10].
probabilities. A game with a randomized (or mixed) strategy is
one in which all of a player’s choices add up to a value of one.
A mixed strategy is comprised of possible actions and an
associated probability. Any mixed strategy that guarantees an
expected reward at least equal to the value of the game is an
optimal strategy [24]. Our model’s possible choices of
movement are:

x1 = probability that blue player chooses north

x2 = probability that blue player chooses north east
x3 = probability that blue player chooses east
x4 = probability that blue player chooses south east
x5 = probability that blue player chooses south
x6 = probability that blue player chooses south west
x7 = probability that blue player chooses west
x8 = probability that blue player chooses north west

Our example can be considered a game in which the red

player, or nature, has already placed obstacles due to destruction
in the path of the blue player and now it is the blue player’s Fig 7.a Optimal Path to Drop Off Center b. Damage Areas
responsibility to navigate through the scene as shown in Figure 6.
Our solution is dynamic since decisions of movement are made at Our solution uses a Markov processes since the current
every point on the grid along the path. The solution for the blue state depends only on a finite history of previous states. For the
player’s mixed strategy in terms of probabilities 𝑥 = first-order Markov process we have:
(𝑛, 𝑛𝑒, 𝑒, 𝑠𝑒, 𝑠, 𝑠𝑤, 𝑤, 𝑛𝑤). We can solve this linear program
using several techniques such as the simplex method, the dual p(rk|{r0, r1, ..., rk−1}) = p(rk|rk−1) (3)
simplex method, or the artificial basis technique. To show how
our system could work, we utilized a computer with LINGO where r is the pixel to be labeled along the traversed path and k
software to solve for the mixed strategies as well as the value of is time. In a Markov Chain we define transition probabilities as
the game. We apply the reward matrix to every grid point value the probability that the system is in state i at time k when in
from start to end. The path from the south is in yellow, in Fig 7a. state j at time k-1 [21]. In our Hidden Markov Model (HMM)
problems we have modeled our transition probabilities as of problems. However, game theory offers important insights
calculated from the Nash Equilibrium. and demonstrates superiority of cooperation over competition.
During the traversal along dynamic path, if a repeat of path is Game theory models the heuristics people use in managing
encountered, then we use next highest mixed strategy probability their conflicts and helps to explain why rational decisions often
direction in order to prevent an infinite loop from occurring. Our miss opportunities for mutual gain [16].
current solution only looks one step ahead. We also have Imperfect information may still be useful to help make
accounted for a rescaling of the weights in the reward matrix if the decisions. Opponent modeling works by observing the
algorithm gets stuck in a loop. opponent’s actions and building a model by combining
Our Markov Decision Process is a sequential decision information from a pre-computed equilibrium strategy with the
problem defined by a state set S, and an action A. The observations [5]. Previous work performed in the community
efficiency, for the blue player, of the algorithms is determined includes computing robust optimization equilibrium by
using the following formula [18]: methods analogous to those for identifying Nash Equilibrium
of a finite game with complete information [1].
Efficiency =
𝐿𝑒𝑛𝑔𝑡ℎ 𝑜𝑓 𝑓𝑖𝑛𝑎𝑙 𝑝𝑎𝑡ℎ
x 100 (4) There is much attention given to simultaneous-move, one-
𝑁𝑢𝑚𝑏𝑒𝑟 𝑜𝑓 𝐶𝑒𝑙𝑙𝑠 𝑆𝑒𝑎𝑟𝑐ℎ𝑒𝑑 shot, normal form games with complete information. Each
player or agent has a private payoff known only to that agent.
Recent advances in self-supervised learning have enabled The payoff to an agent x is not only a function of all the
very long-range visual detection of obstacles and pathways (to agents’ actions (as in the usual complete information game) but
100 hundred meters or more). Unfortunately, the category and also of the realized private-type of agent x. The type of an
range of regions at such large distances come with a agent may be discrete or continuous. Each agent’s realized
considerable amount of uncertainty. A mapping and planning type is chosen independently from some commonly known
system that accurately represents range and category distribution over types, and the payoff matrices for the agents
uncertainties, and accumulates the evidence from multiple are also common knowledge. These games have incomplete
frames in a principled way are desired [19]. information because each agent must choose its strategy, i.e.,
C. Bayesian Decision Making its probability distribution over its actions, without knowing
the realized types of all the other agents [20].
When a natural disaster occurs, one of the first decisions Harsanyi proposed a method for transforming uncertainty
that a country or player makes is whether or not to participate over the strategy sets of players into uncertainty over their
or enter into the game. This is called an entry game. Another payoffs. The transformation appears to rely on an assumption
question which occurs is the how much commitment in terms that the players are rational. Without a common belief of
of contribution is desired. Models decisions may be based on rationality, such implications are not necessarily maintained
the toughness of the incumbent. under a Harsanyi transformation. Under the belief system
A large class of sequential decision making problems model, such implications can be maintained in the absence of
under uncertainty with multiple competing decision common belief of rationality [6].
makers/agents can be modeled as stochastic games. Non- Generally, players may not possess full information about
cooperative games can be solved in which each decision maker their opponents. In particular, players may possess private
makes his own decision independently and each has an information that others should take into account when forming
individual payoff function. In stochastic games, the expectations about how a player would behave. To analyze
environment is non-stationary and each agent’s payoff is these interesting situations, a class of games with incomplete
affected by joint decisions of all agents, which results in the information was created as use case scenarios (i.e., games
conflict of interest among decision makers [12]. Nash where at least one player is uncertain about another player’s
Equilibrium game theory considers the effect of a player’s payoff function) which are the analogue of the normal form
decision on other decision makers [13]. games with complete information similar to Bayesian games
Bayesian game is an interactive decision situation (static games of incomplete information) [22].
involving several decision makers (players) in which each Several studies provide discussion and attempts to
player has beliefs about (i. e. assigns probability distribution integrate and validate usefulness of the application of game
to) the payoff relevant parameters and the beliefs of the other theory models. The strategy action game is not only applicable
players. It is convenient to think of a state of nature as a full in the field of commercial negotiation; subsequent research can
description of a ‘game-form’ (actions and payoff functions). extend further into the fields of education, marketing, finance,
Type also known as state of mind, is a full description of risk management, and society. The competition and
player’s beliefs (about the state of nature), beliefs about beliefs cooperation relationship between manufacturer and distributor
of the other players, beliefs about the beliefs about his beliefs, in other applications are delicate, allowing room for other
etc. ad infinitum. State of the world is a state of nature and a methods besides strategy action game, such as series
list of the states of mind of all players [25]. bargaining game and mean difference. Studies have been
Game theory, as a model of conflict, suffers from several performed on the analysis aiming at the strategy application,
limitations. Players are assumed to always maximize their and intervention into the negotiation harmonization with the
outcomes. Not all of the payoffs or situations can be quantified manufacturer or distributor. On one hand, it insists on an
in a reward matrix. Game theory is not applicable to all types
objective observation attitude; on the other, it may also
produce the deviation of unscrambling the behavior of game P(¬U|i, ¬i) =
P(i,¬i|¬U)P(¬U)

participants subjectively [14]. In our example, let: P(i,¬i|¬U)P(¬U) + P(i,¬i |U)P(U)

(1−p)pq
P(¬U|i, ¬i) =
(1−𝑝)𝑝𝑞 + 𝑟(1−𝑟)(1−𝑞)
P(U) = Probability of United States Giving
P(¬U) = Probability of United States Not Giving Let n = the number of countries. A country will give if [7]:
P(S) = Probability of Saudi Arabia Giving
P(¬S) = Probability of Saudi Arabia Not Giving (1−𝑝)𝑝𝑛−1 q
z ≥
(1−𝑝)𝑝𝑛−1 q+ 𝑟(1−𝑟)𝑛−1 (1−𝑞)
P(S|U)(P(U) 1
P(U|S) = z ≥ (5)
P(S) 1+[𝑟/(1−𝑝][(1−𝑟)/ 𝑝)𝑛−1 ]([1−𝑞)/𝑞]
P(S|U)P(U)
P(U|S) =
P(S|U) P(U) + P(S|¬U)P(¬U)
Given that p > 1-r, and p > 1/2 > 1 – q, the denominator
approaches 1 as n increases. Therefore, as n increases, if
Using event space, j = {give, not give},
P(S|U)P(U)
countries are motivated even slightly to give when there is no
P(U|S) = need without future benefit (type II error), then a country who
∑j P(S|Uj) P(Uj)
interprets a piece of information as motivation for giving will
Suppose a country receives an update about a situation, i, nevertheless not give. Our model indicates that a country is
and interprets the information as motivation for giving aid, the less concerned about giving when there are multiple n givers.
probability of a country not giving when there is a desire for In this Nash Equilibrium the probability of a country not
future cooperation with the Philippines is: giving increases as n increases. Figure 8 shows a plot of z
values with decreasing values for z as n increases using
P(i|¬U)P(¬U) Equation 5. As q decreases, the value of z increases.
P(¬U|i) =
P(i|¬U)P(¬U) + P(i|U)P(U)

• Let p = probability of not giving with no desire for future

cooperation
• Let r = probability of giving with desire for future
cooperation
• Let q = probability of not giving (historical)

(1−p)q
P(¬U|i) =
(1−p)q + r(1−q)

The country’s payoffs are:

• 0, -if not give and do not desire help in future
-if give and desire help in future Fig 8. Plot of z values vs. n players
• -z, -if not give and desire help in future
• -(1-z), -if give and do not need help in future Figure 9 shows the reward matrix of aid provided for
• z is cost of missing opportunity for giving (type I error), countries competing for good will towards Philippines. Solving
• 1 - z -cost of giving without future benefit (type II error). the reward matrix using a Bayes-Nash solver shows that the
United States is best choice for Philippines as a national
We model that a country will give if (1-z)P(¬U|i) ≤ z(1- strategic ally.
P(¬U|i)):
(1−p)q
z ≥ P(¬U|i) =
(1−𝑝)𝑞 + 𝑟(1−𝑞)

We model that a country will not give if:

pq
z ≤ P(¬U|i) =
𝑝𝑞 + (1−𝑟)(1−𝑞)

A country optimally acts according to interpretation of

situational awareness information as motivation for giving.

pq (1−p)q
≤z ≤
𝑝𝑞 + (1−𝑟)(1−𝑞) (1−𝑝)𝑞 + 𝑟(1−𝑞)

If another country is not motivated to give, then a country may Fig 9. Adversarial Planning
not be as motivated to give (or give as much):
 V. CONCLUSION [7] Huang, H., “Introduction to Game Theory Lecture Note 7:
Bayesian Games”, University of California, Merced, Fall 2011.
No decision is ever 100% correct; however, understanding [8]http://world.time.com/2013/11/18/typhoon-haiyan-u-s-pledges-
the effects of algorithmic decisions based upon multiple more-aid-to-philippines-recovery/
variables, attributes, or factors and strategies with probability [9]http://www.innocentive.com/blog/2013/11/25/the-impact-of-
assignments can increase the probability for the best decision crowdsourcing-on-typhoon-haiyan-response/
for a particular situation or event. GlobalSite can perform open [10]http://www.21stcentech.com/communications-update-
source discovery and data mining activities to parse crowdsourcing-relief-agencies-typhoon-haiyan/
information found from disparate, non-obvious, and previously [11] Kearns, Michael, Michael L. Littman, and Satinder Singh.
"Graphical models for game theory." Proceedings of the Seventeenth
unknown data sources and allows for the user to dial the conference on Uncertainty in artificial intelligence. Morgan
weighting factors based upon their knowledge or expertise. Kaufmann Publishers Inc., 2001.
We discussed a method for modeling asset management [12] Li, J., Learning average reward irreducible stochastic games:
with limited resources. We realize that solution presented is analysis and applications, PhD thesis, University of South Florida,
Department of Industrial and Mgmt System Engineering, 2003.
only a guide and is not intended to replace the human brain in
[13] Liggins, Hall, Llinas, “Handbook of Multisensor Data Fusion,
decision making. We offer a user assisted means of Theory, and Practice”, 2nd Edition, 2009.
prioritization to make agent and resources more effective. [14] Lo, Chih-Yao, Yu-Teng Chang, “Strategic analysis and model
Automated game theory is promising for automatically solving construction on conflict resolution with motion game theory,” Journal
real world strategies and helps the security analyst make of Information and Organizational Sciences, 34.1, 2010, 117–132.
optimal decisions for target tracking and detection activities. [15] Nash, John (1951) "Non-Cooperative Games" The Annals of
Automated processing techniques are needed to augment Mathematics 54(2):286-295.
tactical intelligence-analysis capabilities by identifying and [16] Picard, Rick, Todd Graves, and Jane Booker, “Stability modeling
and game-theoretic considerations,” Los Alamos National Laboratory
recognizing features of obstruction. Technical Report, 1999.
We have identified a mathematical application using linear [17] Rahmes, M., Wilder, K., Yates, H., Fox, K., “Near Real Time
programming optimization. Our solution provides the ability to Discovery and Conversion of Open Source Information to a Reward
populate a reward matrix from remotely sensed data. We Matrix”, WMSCI 2013, 12 July 2013.
calculate optimal strategies for path optimization which [18] Razavian, Adam A., and Junping Sun. "Cognitive based adaptive
path planning algorithm for autonomous robotic vehicles." In
increases likelihood of best decision available using game SoutheastCon, 2005. Proceedings. IEEE, pp. 153-160. IEEE, 2005.
theory in a constant sum game. We combine a number of [19] Sermanet, Pierre, Raia Hadsell, Marco Scoffier, Urs Muller, and
technologies for data fusion/ visualization. Our solution is a Yann LeCun. "Mapping and planning under uncertainty in mobile
multi-use application: course of action (COA) planning, robots with long-range perception." In Intelligent Robots and Systems,
strategies, resource management, risk assessment, etc. 2008. IROS 2008. IEEE/RSJ International Conference on, pp. 2525-
2530. IEEE, 2008.
Automated processing techniques are needed to augment
[20] Singh, S. P., V. Soni, M. P. Wellman, “Computing approximate
tactical intelligence-analysis capabilities by identifying and bayes-nash equilibria in tree-games of incomplete information”;
recognizing patterns, weighting them appropriately, providing Proceedings of 5th ACM Conference on Electronic Commerce, 2004,
near real time objective decisions where the user can interact pp. 81–90.
with the information based upon their experiences and [21] Skoglar, Per. "UAV path and sensor planning methods for
multiple ground target search and tracking-A literature survey."
knowledge base. GlobalSite is a probabilistic decision solution Department of Electrical Engineering, Linköping University, Tech.
which allows for users to interact with information in near real Rep (2007).
time using game theory to provide a reward matrix of the best [22] Slantchev, B., “Game Theory: Static and dynamic games of
possible outcomes. incomplete information,” Dept of Political Science, Univ San Diego,
May 15, 2008.
REFERENCES [23] Strode, Christopher. "Optimising multistatic sensor locations
[1] Aghassi, M., and D.Bertsimas, “Robust game theory,” using path planning and game theory." Computational Intelligence for
Mathematical Programing, Ser. B, vol. 107, 2006, pp. 231–273. Security and Defense Applications (CISDA), 2011 IEEE Symposium
on. IEEE, 2011.
[2] Brandenburger, Adam. "Cooperative Game Theory." Teaching
Materials at New York University (2007). [24] Wayne Winston, Operations Research Applications and
Algorithms 4th. Edition, 2003.
[3] Campbell, A. T., Eisenman, S. B., Lane, N. D., Miluzzo, E.,
Peterson, R. A., Lu, H., ... & Ahn, G. S. (2008). “The rise of people- [25] Zamir, Shmuel. Bayesian games: Games with incomplete
centric sensing”. Internet Computing, IEEE, 12(4), 12-21. information. Springer New York, 2009.
[4] Diplaris, S., Papadopoulos, S., Kompatsiaris, I., Goker, A.,
Macfarlane, A., Spangenberg, J., ... & Klusch, M. (2012, April).
SocialSensor: sensing user generated input for improved media
discovery and experience. In Proceedings of 21st intl conference
companion on World Wide Web (pp. 243-246). ACM.
[5] Ganzfried, S., and T. Sandholm, “Game theory-based opponent
modeling in large imperfect-information games,” International
Conference on Autonomous Agents and Multi-Agent Systems
(AAMAS), 2011.
[6] Hu, Hong, and Harborne Stuart, Jr., “An epistemic analysis of the
Harsanyi transformation,” Intl Journal Game Theory, 2002, pp. 517–
525.