48 AI MAGAZINE
( 1) by changing sensory input (information seek-
ing action selection), ( 2) by changing predictions
of sensor inputs (perceptions, beliefs), and ( 3) by
changing the model of the decision maker’s team
structure and coordination (learning). Variational
free energy is a function of sensory outcomes
(data) and probability density over their (hidden)
causes (true world states or context). This function
is an upper bound on surprise, a negative log of
model evidence representing the difference be-
tween agent predictions about sensory inputs, and
observations or data encountered. Indeed, differ-
ences between variational free energy and surprise
is Kullback–Leibler divergence between agent be-
liefs about context (called the recognition density)
and the joint density of context and data given the
agent model (the generative density). Because the
long-term average of surprise is entropy, an agent
acting to minimize free energy will implicitly place
an upper bound on the entropy of outcomes or
sensory states sampled. Consequently, the free en-
ergy principle provides a mathematical foundation
to explain how agents maintain order by restrict-
ing themselves to a limited number of perceived
high-probability and high-utility (context, action)
pairs. This restriction gives a formal mechanism
to inference and decision making, where multiple
agents operate autonomously, coordinate among
themselves, and resist disorder (Levchuk et al. 2018).
Context inference compares model predictions
of outcomes with expected (desired) outcomes;
these deviations form the basis for nonnormalcy
detection, active learning, and subsequently predictions based on updated context model. Because context inference is a maximum a posteriori
estimation problem on a DHBN, we convert the
DHBN model into a factor graph of data variables,
context factors, and decision makers or agents.
Algorithms for dynamic context inference using
factor graphs of data variables and context factors
(assuming each factor is associated with an agent)
include generalized belief propagation algorithms,
decomposition algorithms, and clustering approaches.
Generalized belief propagation algorithms (Yedidia,
Freeman, and Weiss 2005) include a combination of
coordinate descent, Lagrangian relaxation and Viterbi
decoding algorithms developed for coupled HMMs
(Zhang et al. 2013), and semisupervised (active
learning–based) clustering algorithms. Because exact
computation of posterior context distribution is NP-hard, an approximate solution is produced by the
max-product belief propagation algorithm, which
Workload
ASW Mine_Warfare ISR AIr
Sea_Surface
Underwater
Airborne
Surface
Assets
Track
Search
Detect
Classify
Knowledge
Skills
Role
Sub–surface
Risk_Propensity
Time_Stress
Human
Threats/Tasks
Mission Environment
MEAT–H +
Figure 3. Representing Contextual Elements.
Contextual elements (MEAT-H) are represented in hybrid teams for maritime decision making via Protégé (Mishra et al. 2017b). Solid
arrows represent the hierarchy, whereas dashed arrows denote relationships between elements. In this figure, mission elements involve
antisubmarine warfare, mine warfare or intelligence, surveillance, and reconnaissance operations.
