-We consider the problem of real-time, proactive decision making for dynamic and time-critical decision-events where the choices made for multiple, individual decisions over time determine the final decision outcome of an event. We posit that the quality of such individual decisions can be significantly improved if human decision makers are provided with decision aids in the form of dynamically updated information and dependencies between the different decision variables, and the humans affecting those decision variables. In this position paper, we propose the CONRAD (CONtext aware Real-time Adaptive Decision making) system that uses computational techniques from large scale network analysis and game theory-based distributed information aggregation to develop such decision aids. CONRAD's functionalities are implemented through three subsystems - a decision making subsystem that updates and mathematically combines information from different decision variables to predict the outcome of the decision event, a decision assessment subsystem that uses the currently predicted decision outcome to estimate the future decision trajectory and recommends information collection-related actions to the human decision maker, and, a network analysis subsystem that uses those recommended actions to dynamically update the dependencies and correlations between events and people influencing the decision variables. To the best of our knowledge, our work is one of the first attempts towards combining dynamic decision updates and using the predicted decision trajectory as a proactive feedback mechanism to dynamically update the correlations between decision variables so that human decision makers can make more strategicallyinformed and well-aligned decisions towards the desired outcome of decision events.
I. INTRODUCTION
Modern decision making scenarios are characterized by
large amounts of data and information that arrive dynamically,
over a short period of time, from multiple sources. Processing
this data in a time-critical manner to make accurate decisions
is an overwhelming task for human decision makers. Over
the past few decades several decision making solutions have
been proposed to aid human decision makers with tools such
as intelligent or automated software that use computational
methods and mathematical models of human cognitive
processes to make sophisticated decisions on behalf of humans[
To address this problem, we present the framework of a context-aware, real-time, decision making system called the CONRAD (CONtext aware Real-time Adaptive Decision making) system that focuses on enabling and enhancing the capabilities of human decision makers by developing proactive decision-aids for making high-accuracy, time-critical decisions in complex, data- and information-rich environments. The central research problem that CONRAD proposes to answer is as follows: Given a set of decision variables in the current decision context along with a set of data sources from which the decision variables can be derived and/or calculated and updated, what is a suitable set of techniques for (i) extracting relevant information, (ii) then using that information to update, correlate and aggregate the decision variables dynamically, and finally , (iii) assessing the quality of the aggregated decision outcome (prediction), so that the divergence between the aggregated decision outcome (predictions) and the desired decision outcome is successively minimized?
To address this question, we propose to represent a decision event as a collection of decision variables that are affected by the data from the environment. The system dynamically determines the inter-dependencies between these decision variables and also periodically updates them into an aggregated decision outcome (prediction). This aggregated decision outcome is then evaluated with respect to the desired decision outcome, and, depending on the deviation between the actual and desired decision outcomes, actions are recommended to collect additional data/information and discover new data correlations. This information is then used to update the decision variables autonomously and proactively - so that the quality of the aggregated decision outcome successively converges towards the desired outcome. We plan to realize the aforementioned functionalities in CONRAD using three subsytems that are summarized below: (1) Decision Making Subsystem: The decision making subsystem uses a prediction market-based information aggregation mechanism to update and mathematically combine or aggredec1 dec2
Sliding
Autonomy Goal-directed Decision Making Component Updates and aggregates decision variables, predicts decision outcome
Pred.
Dec.
Out-come
Update dec. vars with new informa!on and new correla!ons
Past/Similar Decisions Decision Assessment
Component Calculates devia!on of predicted and desired decision outcomes, suggests ac!ons based on relevance to domain/past decisions
Devia!on, Sugg. ac!on
Dynamic Info Extrac!on & Valua!on
Component Iden!fies cri!cal/info. deficient dec. vars, extracts targeted new info, discovers new info correla!ons, etc. gate information from different decision variables and predict the outcome of the decision event. (2) Decision Assessment Subsystem: The decision assessment subsystem uses the currently predicted decision outcome from the decision making subsystem along with relevant domain knowledge from past decisions made in similar domains to predict the decision trajectory and recommends information collection-related actions to the human decision maker. Machine learning and AI-based planning techniques are used to implement the functionalities of the decision assessment subsystem. (3) Dynamic Information Extraction and Valuation Subsystem: The dynamic information extraction and valuation subsystem uses the actions recommended by the decision assessment subsystem to model and dynamically update the dependencies and correlations between events and people that influence the decision variables using metrics and techniques from large scale network analysis. The different components of CONRAD and their main functionalities are given in Figure 1 and discussed in the following sections.
The main research question addressed in CONRAD’s goal
directed decision making subsystem is how to design a suitable
set of computation techniques to dynamically update the
different decision variables in the current decision context and
combine or aggregate them into a single, global decision outcome.
The decision variables are extracted from the environment’s
information by CONRAD’s information extraction component,
discussed in Section IV. We propose to perform the update
and aggregation of the individual decision variables using
an information aggregation technique inspired by prediction
markets. Prediction market based information aggregation [
Dynamic decision variable updates. To enable dynamic updates of the decision variables, we associate each decit with a decision making (or belief update) agent ai; ai is responsible for updating the value of decit at time step t. Agent ai performs this update using the following belief update formula: decit = beli(decit−1, dect−−i1, AggDect−1), where beli(.) is the belief update function used by agent i, decit−1 is the value of deci during time step t−1, dect−−i1 is the set of decision variables from time step t−1, excluding decit−1 itself, that are correlated with decit during time step t and AggDect−1 is the value of the aggregated decision outcome during time step t − 1. The decision maker agent also ensures that decision variables that have already converged to their optimal or best value are not updated. The decision maker agent uses the intelligence, from reviewing the current context, to identify only those decision variables that need updating.
Aggregating decision variables. At the next step, the
individual decision variables are combined into an aggregate
or predicted decision outcome by the aggregator agent. A
market-based aggregation mechanism provides a suitable way
to combine information from multiple sources (e.g., multiple
decision variables updated by the decision maker agents) into
a single aggregated decision outcome value using a technique
called a scoring rule [
III.
DECISION ASSESSMENT
The objective of CONRAD’s decision assessment component is to determine how well the current aggregated (predicted) decision outcome is aligned with the desired decision outcome and to recommend actions related to future information collections that could potentially improve the convergence of the predicted decision outcome towards the decision outcome. A schematic of the decision assessment component is shown in Figure 3. Because we have represented decision outcomes as probability distributions (belief values), statistical divergence metrics such as the Kullback-Leibler (KL) divergence can be used to predict the future decision trajectory - Some Decision Variables & Correla!ons deci t-1 : … Decision !me deci t : .
: Updated Decision Vars.
… Informa!on Aggregator
Agent Aggrega!on Mechanism Dec. Agent Scoring (Payment)
Rule …
Agent Rewards/Penal!es
Predic!on Market-based Mechanism for Informa!on Aggrega!on
well-known decision trajectories can be constructed from past
decisions and then Bayesian inference can be used to classify
the current decision trajectory into one of the trajectory types.
The historical aggregated decisions can be further refined
with domain knowledge to reflect the changes in the situation
since the decisions were aggregated. The decision assessment
subsystem also suggests actions related to future information
collection to the human decision maker and to CONRAD’s
information extraction/evaluation component using AI-based
planning techniques such as MDPs and POMDPs [
INFORMATION EXTRACTION AND VALUATION
The key to efficient decision making is to ensure that the available information is dynamically updated and important correlations in data are accurately captured. To achieve these objectives, CONRAD will perform the following operations in real time;
Extract Decision Variables from Raw Data. The data
extraction tool of CONRAD extracts data from different
heterogeneous and potentially changing sources and filters
decision variables - id of the data creator, the data creation
time, and a list of key fields such as demography, topic
of discussion, etc. We will use Semantic Technology and
represent the list of fields through an ontology based language
such as OWL. Our goal is to create a database similar to
DBpedia (dbpedia.org/About) that will allow users to submit
queries with multiple conditions and identify entities that fulfill
those queries. The correlations between the decision variables
are modelled as networks (or graphs). The vertices in the
network represent the entities and the edges represent the
correlations. Using this collected data, the information
component performs the following subtasks: (a) Creating Multilevel
Networks. A network is created from the processed data as
follows - one field in the dataset is identified as the entity
variable and other selected field(s) as the relation variable(s).
Each vertex of the network represents a unique instance of
the entity variable (here each entry is the name) and two
entities are connected if they satisfy certain relations between
the relation variables (for example, ids with age difference of
five years or less are connected). The connectivity patterns
of the networks can with the time stamp changes. Networks
based on the same entity variable can be further combined
to a multilevel network. This enables us to unearth obscure
information that is not immediately relevant from only one
database. (b) Real Time Analysis of Networks. The analysis of
the networks provides insights to the characteristics of the data.
Some of the common analysis objectives in CONRAD include
(i) detecting communities to identify tightly connected groups
of vertices [
CONRAD performs network analysis operations at three
levels, as illustrated in Figure 4. The first is the horizontal
level that analyzes each entity network. The second is the
longitudinal level where the analysis is conducted across levels
(the networks at each level have the same entity variable,
but the relational variable(s) and therefore the structure is
different). The third level is the temporal level where we
track the changes to the network structure across different
time steps [
Identifying Critical Decision Variables. Identify
important decision variables that can predict future events will
enable users to maintain the correct decision trajectory. The
critical variables are the ones whose corresponding networks
guarantee that the analysis results are accurate under various
perturbations to the data and are sensitive to changes in the
data. CONRAD evaluates the reliability of the network models
based on well-posedness and the sensitivity with respect to the
analysis objective. To the best of our knowledge, our work
is one of the first instances that a network analysis toolkit
will include a component to compute the accuracy of the
data. Well-posedness is a measure of whether the analysis
objective, is feasible for a given network. To compute
wellposedness of a network, CONRAD computes the number of
solutions that the network has for a given analysis function.
For community detection, this can be computed by changing
the vertex ordering, and then taking the consensus of the
communities obtained at each ordering to find the well-posed
subgraph [
Integrating Decision Making and Data Extraction. The final objective of the information extraction and valuation component in CONRAD is to use the recommendations from the decision assessment algorithm to update the information networks. Based on the recommended actions, the information component tries to extract ’meaningful information’ from ’raw data’. The main operations of this process are (i) improving the data gathering mechanism, (ii) improving the quality of the networks and (iii) improving accuracy of the analysis. Data Gathering. The data gathering operation can be improved by adding more varied sources of information. For example, we can enrich information about possible disasters, by including information of past hurricanes and earthquakes, in addition to tracking the current disaster through news sources, and social network sites.
Adaptive Refining of Data Data is generally gathered
’wholesale’, without specifically considering the subsequent use of
the information. In the network modeling stage, the system
refines the data by filtering the initial network based on certain
combinatorial properties. For example if the agents’ focus is on
finding clusters of similar entities, then a chordal graph based
filtering that will retain only the tightly connected components
in the network is used. Conversely if the decisions are to be
based on centrality metrics, then filtering to reduce the low
weight edges is more effective [
V.
CONCLUSION
In this position paper, we have proposed CONRAD, a realtime, proactive decision aiding tool that leverages the advantages of game theory, machine learning and network analysis. Each of the individual components proposed for CONRAD have been shown to be successful in their respective domains and we posit that combining them will further enhance the decision making capabilities.