Spatio-Temporal Reasoning within a Neural Network framework for Intelligent Physical Systems

Full text

Spatio-Temporal Reasoning within a Neural Network
framework for Intelligent Physical Systems
Sathish A.P. Kumar
Department of Computing
Sciences, College of Science
Coastal Carolina University
Conway, SC, USA 29528
skumar@coastal.edu
Michael A Brown
Research & Development,
Intelligent Systems Department
South West Research Institute
San Antonio, TX, USA 78228
michael.brown@swri.org
Abstract—Existing functionality for intelligent physical
systems (IPS), such as autonomous vehicles (AV), generally lacks
the ability to reason and evaluate the environment and to learn
from other intelligent agents in an autonomous fashion. Such
capabilities for IPS is required for scenarios where an human
intervention is unlikely to be available and robust long-term
autonomous operation is necessary in potentially dynamic
environments. To address these issues, the IPS will then need to
reason about the interactions with these items through time and
space. Incorporating spatio-temporal reasoning into the IPS will
provide the capability to understand these interactions. This
paper describes our proposed neural network framework that
incorporates spatio-temporal reasoning for IPS. The preliminary
experimental results addressing research challenges related to
spatio-temporal reasoning within neural network framework for
IPS are promising.
Keywords— Spatio-Temporal Reasoning, convolution neural
networks, Automated Vehicles
I. I
NTRODUCTION AND
B
ACKGROUND
Intelligent Physical Systems such as Automated Vehicles
has the potential to impact the society with far-reaching
applications and implications beyond all current expectations
[34-38]. For example, automated vehicles can be applied for
remote planetary exploration, emergency management and
disaster rescue operations [39-41]. IPS in these scenarios
would be able to reason from a spatio-temporal perspective
interacting with, and perhaps even manipulating, the
environment [42]. In a typical IPS, the contact with human and
even the option of direct remote control is very limited [43]. So
it is extremely important for the IPS to reason the environment
from a spatio-temporal perspective. The concept of flexibility
and spatio-temporal awareness is further important if they need
to exhibit operational resilience for extended periods of time
and at the same time need to respond to unanticipated scenarios
without human intervention in a remote area [44]. When the
multiple IPS collaborate together in the environment, they can
share their contextual knowledge (including spatio-temporal
information) directly and observe other vehicle’s behaviors
[45]. It is necessary for intelligent agents based IPS to evaluate
the environment they operate in to decide on the right action
and execute in an autonomous fashion. However, this is not an
easy task as many factors such as uncertain knowledge
including from a spatio-temporal perspective will affect the
IPS to deduce the environment they are operating with.
However, it is assumed that the IPS should be able to
efficiently understand the environment they are interacting
with over a period of time. Hence a framework is essential for
the agent to perceive the environment as well as learn from the
other agents operating in the environment [46].
The sustained research and development efforts for IPSs,
could result in a robust AVs, which has the potential to be an
economic disruptor in the near future [47-49]. In order to reach
the level of robustness that is required, the AVs should be able
to handle unique inputs or changes to the environment in which
they operate. However, the current methods in AVs are unable
to account for changes from spatio-temporal perspective [50].
In this paper, we propose research objectives to address this
limitation through the incorporation of spatial-temporal
reasoning within an IPS, and exploration of the environment
through experimental learning. Figure 2 describes the proposed
architecture, which incorporates spatio-temporal reasoning
within a NN framework.
Contributions: Through our approach, intelligent agent
based IPS can adjust their behavior and learn new behavior
based on the experiences through observation and interactions
with other IPS. Proposed neural network framework
incorporating spatio-temporal reasoning is as shown in Figure
2, will endow the intelligent agent with reflective capabilities
to respond to its environment through introspection and
observation, which allows it to dynamically expand its
operational design domain. The proposed framework, as shown
in Figure 2, would also enable the IPS to learn and adapt to
improve their behavior over time, through their knowledge and
experience gained by observing or interacting with the
environment. From a broader societal impact perspective,
proposed solution can also be applied in space exploration,
search and rescue operation, and in remote emergency response
in a hazardous environment, potentially saving lives.
Rest of the paper is organized as follows. Section 2
describes the related work. Section 3 describes the research
objectives required for incorporating spatio-temporal reasoning
within neural network framework for IPS. Section 4 describes
the methodology and architecture of spatio-temporal reasoning
within the neural network framework. Section 5 describes the
274
978-1-5386-9276-9/18/$31.00 c
2018 IEEE

preliminary experimental results with respect to research
objectives outlined in Section 3. Section 6 describes the future
work and finally section 7 presents the summary and
conclusion
.
.
II. R
ELATED
W
ORK
In a related work with respect to Spatio-temporal reasoning
within a neural network, Z. Kira et al., [1] specifically combine
the strengths of CNN with temporal models, in the form of
Monte Carlo Localization methods, in order to perform higher
level reasoning tasks for everyday environments. They also
discuss on additional potential for higher-level spatio-temporal
environment understanding tasks such as inference, prediction,
and generalization across environments. Neural networks using
spatio-temporal reasoning have been used for applications
such as prediction in driving videos [2], neural network-based
reasoning over natural language sentences [3], analyzing user
behaviors in predicting their next location [4], prediction of
Futures Contract Prices of White Maize in South Africa [5],
contributions of rainfalls to flash floods [9], short term
prediction of traffic flow [10], air temperature prediction [11],
public transport services [12], for personalized decision
support system [15], predicting resource usage [17], human
action recognition [18], and intelligent environmental decision
support systems [20]. In another related work, Kasabov et al.,
describes a new type of evolving connectionist systems
(ECOS) called evolving spatio-temporal data machines based
on neuromorphic, brain-like information processing principles
[6, 7, 16, 19]. Guesgen, and Marsland propose spatio-temporal
reasoning and context awareness in the context of smart homes
[8]. A stochastic graph based framework was suggested by
Xiaong et al. for a robot to understand tasks from human
demonstrations and perform them with feedback control. Their
framework unifies both knowledge representation and action
planning in the same hierarchical data structure, allowing a
robot to expand its spatial, temporal, and causal knowledge at
varying levels of abstraction [13]. In another related work, Jain
et al [14], propose an approach for combining the power of
high-level spatio-temporal graphs and sequence learning
success of Recurrent NNs. Ohn-Bar and Trivedi studied an
idea of object relevance, as measured in a spatio-temporal
context of driving a vehicle. In this they applied spatio-
temporal object and scene cues for object importance
classification [58].
As per our literature review, there is no related work on
augmenting observed learning and spatio-temporal learning in
an IPS environment. While it is not directly applied in a
physical systems environment, experimental learning for
intelligent agents is proposed in other areas such as training
salesmen and managers in the context of a retail store or a
larger supermarket [24], operations management [25], text
learning [27], neuroscience [28], digital legal libraries [29],
online learning [30], educational games [31], business
intelligence and consumer intelligence [23] and mobile
learning system [33]. Some of the techniques that have been
applied for learning from experience in multi-agent systems
include decision theoretic intelligent agents [26], leveraging
trust modeling based on direct experience with incentives for
honesty [32], visual question answering [59], video object
segmentation [60] and portable semantic web crawling and
indexing [23].
III. R
ESEARCH
O
BJECTIVES
To address the challenges of realizing IPS such as
automated vehicles and limitations of the related work outlined
in earlier sections, following are the research objectives that the
rest of the paper focuses on.
A. Research Objective 1 (RO1) - Spatio-temporal reasoning
within a neural network enabled IPS
:
In this research objective, our focus is to utilize spatio-
temporal reasoning within a neural network, resulting in neural
networks which model the world more like humans do.
Through the classification of tasks and the context for that task,
spatio-temporal reasoning provides situational awareness to the
intelligent physical system [52]. We hypothesize that current
representations, such as Long-term Recurrent Convolutional
Networks (LRCN) [21], which classify actions and caption
video, suffer from over training because the network designs
do not allow for spatial invariance beyond the pixel level. Due
to this, the utilization of time and space has been suboptimal in
these instances. One noteworthy endeavor to apply these
concepts to robotics is Nvidia's DeepDrive system, which
directly regresses the steering wheel angle based on the
incoming video stream from a forward facing camera [53].
The approach is novel because the system can be trained end-
to-end simply by driving the vehicle, but the front end of the
system is still a classical classification/regression based
architecture. Essentially, the output is a left or right
classification with a magnitude. The network itself lacks the
practical, reasonable considerations and concepts such as
drivable surface and spatial occupancy that make modern
robotic navigation possible.
We propose a projection layer which will project a learned
descriptor of each occupancy voxel into a world model tensor.
This can be used within a neural network to store properties
relevant to interacting with the environment. Properties, such
as static, dynamic, movable, hot, cold, etc., can be processed
with fully convolutional layers in order to make predictions
about the future state of the spatial model based on current
content and history. Because the processing will be done with
Figure 1: IPS with various autonomous capabilities developed by SwRI-165
Military HMMWV
IEEE Symposium Series on Computational Intelligence SSCI 2018 275

fully convolutional approaches, the mappings that are learned
will not depend heavily on viewpoint or require large amounts
of data to account for spatial differences [54]. Success in this
area will reduce training of applicable neural networks by an
order of magnitude in terms of time and size, and identify areas
that undertrained.
B. Research Objective 2 (RO2) - Augment observed
procedures and spatio-temporal analysis
In this research objective, our focus is to augment observed
procedures and spatio-temporal analysis through experimental
learning. Knowledge gained via observation and spatio-
temporal reasoning will be suboptimal [55]. This is due to
several factors, including differences between platforms,
improper communication and identification, and small sample
sizes. In addition, the first systems to encounter a new
environment will not have the benefit of group knowledge.
Experimental learning allows an IPS to simultaneously
perform actions and improve its performance [56]. The system
performs actions that it believes are safe, whether to perform a
novel task or to optimize a task learned via observational or
transference learning. Once the action has been performed, the
vehicle evaluates the environment and determines if it is nearer
or further away from the goal. With each execution of a task,
the system becomes more confident in its ability to perform a
task, which will be considered in a cost model based task
allocation system [57].
IV. S
PATIO
-T
EMPORAL
R
EASONING WITHIN A
N
EURAL
N
ETWORK
F
RAMEWORK
In this section, we describe the needed research efforts to
address the research objectives outlined in earlier sections. In
particular, the research tasks, associated with incorporating
spatio-temporal reasoning in NN framework are described in
this section. To address the research objectives we have
developed architecture, shown in Figure 2, which is based on
Caffe machine learning framework [22]. As show in Figure 2,
observational system (spatio-temporal) data are fed to the
Caffe-based neural network frame. The spatial and temporal
neural network is then constructed to predict future
model.
Following are the detailed research steps related to research
objective 1.
1. Incorporate recurrent neural network into voxel-based neural
network: This task will link a recurrent neural network
structure to each spatial descriptor in order to allow the
network to predict temporal properties based on appearance.
This models the ability of humans to determine if an object is
dynamic, static but movable, or firmly affixed to another
structure in the environment.
2. Incorporate 3D convolutions to learn occupancy
relationships: Building on the output from the first task, it
should be possible to extend the network in the spatial layers
by processing them with convolutions in order to model
relationships between objects. This adds additional complexity
to the model, but it allows the network to learn about
interactions between objects such as a person kicking a ball or
a table holding up a coffee mug.
3. Collect data and train network: This entails allowing the IPS
to observe scenes, which are changing in time. In an end-to-
end fashion, we will link appearance based features to temporal
properties and learn interactions by forcing the network to
predict future occupancy states using the recent
history. Because the training will be done in post process, this
training requires no labeling.
4. Test and validate the network (includes re-training if
necessary): Using withheld data, the goal will be to show that
the network can correctly predict reasonable future states of the
environment. Quantitatively, this can be evaluated with the
loss function used in training, and qualitatively, this means
playing the data back using 3D visualization tools. The goal
will be to show that reasonable predictions are made.
Following are the steps related to research objective 2 of the
proposed framework.
1. Utilize reflective neural network to determine high-value
explorations: Through reflection on the activations of the
internal network layers, exploratory actions that result in
strongly confident network activations can be flagged. These
explorations can be used to encourage behaviors that are
outside of the expected operating parameters, effectively
identifying promising actions that could be performed.
2: Utilize spatio-temporal neural network to assess risk of
action: This step would prevent damage to vehicle or
environment. As exploratory behaviors are prioritized through
step 1, predictions can be generated on the future state, and
risks can be identified. For example, a high value exploration
has been identified that includes physically interacting with an
object in the environment. The spatio-temporal neural network
may predict that the interaction would affect the body of the
IPS as well as the object itself. This risk likelihood could be
factored into the decision of taking action.
3. Develop framework for balancing high-value exploration
and risk: Without taking action with some level of risk, the
operation of the IPS will remain completely in the domain of
manually programmed behaviors. A certain level of risk will
need to be taken in order to provide for novel experiences with
the environment. This framework will allow the network to
Figure 2: Depiction of Spatio-Temporal Reasoning within NN Framework.
276 IEEE Symposium Series on Computational Intelligence SSCI 2018

provide a cost-map that incorporates high value, low risk
behaviors into the potential outcome for behaviors.
4. Integrate experimental learning with knowledge base: As the
IPS performs actions that are beyond the preprogrammed
behavior, additional learning will occur. The inputs that are
gathered through this experimental process will be
incorporated as heavily weighted inputs, and included in the
training efforts, leveraging the advantages from all other RO,
to expand the understanding of the world model of this IPS.
V. E
XPERIMENTAL
R
ESULTS
In our experimentation, we have generated multi-scale
convolution neural networks (CNN) and deployed for image-
based object detection within AV systems. The detection
algorithms are capable of isolating and classifying pedestrians
in real-time with higher accuracy than similar Haar-like feature
based algorithms, as shown in Figure 3.
Fundamentally similar structures are currently being
generated and utilized by our autonomous platforms for
representing the world model as shown in Figure 4. The voxel
structure of the model can be noted in Figure 4. However,
these world models are strictly occupancy grids and do not
leverage the attributes of real-world objects as the proposed
research does. In addition to the world model, cost maps are
generated, as shown in Figure 5, in order to differentiate
between multiple path traversal options. Currently, generating
a cost map is done with hand crafted weighting functions (one
of the many manually programmed efforts that advances the
field but does not address novel environments that future IPS
will encounter). These cost maps are tedious to construct and
not optimized in any way (other than tweaks made manually
over time). The current cost map considers the content of a
voxel or ground cell and its position and maps it to a cost with
a piecewise linear function. Costs are then summed through
columns to produce an initial cost map, which is then refined
with a set of filters. We hypothesize that these cost maps can
be replaced with fully convolutional neural networks that can
direct the navigational selections that are made by the spatio-
temporally aware network.
Through expanding capabilities in path planning, we have
performed preliminary experiments into the area of non-
deterministic path planning. Such planning is similar to
experimental learning in that random paths are included in the
estimation for the path of an AV, providing the capability to
experiment with alternative path options. The path planner that
we developed called as “Maverick” is a hierarchical planner
that is capable of efficiently finding feasible paths over long
distances in unstructured environments while taking into
account steering limits and operating in reverse. The planner
comprises a waypoint planner and a rapidly exploring random
tree (RRT) planner that respects kinodynamic constraints. The
waypoint planner uses a simplified vehicle model to quickly
find a directed graph of waypoints from start to goal, which is
then used to bias sampling and speed up computation in the
RRT planner. The Maverick planner is capable of anytime
planning and continuous re-planning, the combination of which
provides a feasible path quickly and then optimizes the
computed path with any remaining computational cycles as
shown in Figure 6. Structural awareness of the environment is
utilized when available but the planner does not require it.
Using this planner, the IPS will prefer to follow roads or trails
but will navigate off them when necessary.
Figure 5. A world-model generated through sensor fusion and
material classification
Figure 4: A generated cost map with overlays of the
Paths that are planned based on this cost map.
Figure 3: CNN detection of a pedestrian from
forward-looking image sensor.
IEEE Symposium Series on Computational Intelligence SSCI 2018 277

VI. R
ESEARCH
C
HALLENGES AND
F
UTURE
W
ORK
R
ECOMMENDATIONS
In this section we describe the research challenges and
future work recommendations associated with incorporating
spatio-temporal reasoning for neural network framework. One
challenge for Spatio-temporal reasoning within a neural
network objective will be the creation of a layer that represents
physical world concepts. Representing this properly for the
implementation and operating space will be the fundamental
research challenge. In addition, handling sensor occlusion in
training sets will be a challenge that the research team expects
to address through the inference of occupancy in the presence
of occlusion. The initial learning of dynamic objects within a
spatio-temporal neural network presents a ‘chicken and egg’
problem in that the neural network must have seen a dynamic
object to learn its physics, and must have learned its physics to
recognize a dynamic object. We believe that manual training
and testing with dynamic objects on one platform will be
sufficient to allow new platforms to recognize and train on
dynamic objects.
The spatio-temporal reasoning aspect of the developed
neural network allows the system to provide a much more
accurate estimation of future states. In order to validate the
spatio-temporal neural network developed in this paper, data
will be recorded which shows objects undergoing a predictable
motion, such as rolling, falling, or being pushed. At several
points during this motion, the spatio-temporal neural network
will begin predictions which are propagated to the end of the
recorded motion. The deviations of this motion will be
compared to other methods discussed in the literature. To
determine the effectiveness of experimental learning within the
IPS an action will need to be selected and manually
programmed into the system. Once the system has had
sufficient time to execute and optimize the action, the value of
the cost function can be compared to the initial value. This can
be repeated for multiple trials and multiple actions.
VII. C
ONCLUSION
Our preliminary experimental results indicate that spatio-
temporal reasoning can be incorporated into the neural network
framework within the context of intelligent physical systems.
Building on the success of the preliminary experimental
results, we plan to further assess and evaluate the research
challenges and objectives outlined in the section 3 and 4 in a
more realistic AV scenario. Through our approach, intelligent
agent-based IPS can adjust their behavior and learn new
behavior based on the experiences through spatio-temporal
reasoning with other IPS. This research proposed in the paper
will lead to practical guidelines for developing IPS that has the
spatio-temporal reasoning capabilities incorporated into it,
resulting in minimal to no human intervention.
A
CKNOWLEDGMENT
The authors would like to acknowledge the help of
Cameron Mott with respect to the experimentations and the
writing.
R
EFERENCES
[1] Z. Kira, et al, "Leveraging Deep Learning for Spatio-Temporal
Understanding of Everyday Environments", in proc. of IJCAI 2016
Workshop on Deep Learning for Artificial Intelligence, 2016.
[2] E.Ohn-Bar, and M.M.Trivedi, "Are All Objects Equal? Deep Spatio-
Temporal Importance Prediction in Driving Videos", in Pattern
Recognition, 2016.
[3] B. Peng, et al., "Towards Neural Network-based Reasoning", in proc. of
the NIPS Workshop, 2015.
[4] Q. Liu et al., "Predicting the Next Location: A Recurrent Model with
Spatial and Temporal Contexts", in proc. of the Thirtieth AAAI
Conference on Artificial Intelligence (AAAI-16), pp. 194-200.
[5] K. Ayankoya, A. P. Calitz, and J. H. Greyling, "Using Neural Networks
for Predicting Futures Contract Prices of White Maize in South Africa",
in proc. of the Annual ACM Conference of the South African Institute
of Computer Scientists and Information Technologists, 2016.
[6] N. Kasabov, et al. "Evolving spatio-temporal data machines based on the
NeuCube neuromorphic framework: design methodology and selected
applications”, in Neural Networks vol. 78, 2016, pp. 1-14.
[7] N. K. Kasabov, "Evolving connectionist systems for adaptive learning
and knowledge discovery: Trends and directions", in Knowledge-Based
Systems, vol. 80, pp. 24-33, 2015.
[8] H.W. Guesgen, and S. Marsland, “Spatio-temporal reasoning and
context awareness", in Handbook of Ambient Intelligence and Smart
Environments, Springer US, pp. 609-634, 2010.
[9] T. Darras et al., "Identification of spatial and temporal contributions of
rainfalls to flash floods using neural network modelling: case study on
the Lez basin (southern France)", in Hydrol. Earth Syst. Sci., vol. 19, pp.
4397–4410, 2015
[10] V. J. Hodge, et al., “Short-Term Prediction of Traffic Flow Using a
Binary Neural Network”, in Neural Computing and Applications, vol.
25.no. 7-8, pp. 1639-1655, 2014.
[11] D. Deligiorgi1, et al., “Artificial Neural Network based Methodologies
for the Spatial and Temporal Estimation of Air Temperature” in proc. of
ICPRAM 2013 - International Conference on Pattern Recognition
Applications and Methods, pp. 669-678.
[12] M. Seredynski, et al., "Towards novel public transport services via real-
time optimization of demand and supply with traveler incentivisation",
in proc. of the 23rd ITS World Congress, Melbourne, Australia, 10–14,
October 2016.
[13] Xiong et al., "Robot Learning with a Spatial, Temporal, and Causal
And-Or Graph", in proc. of the International Conference on Robotics
and Automation (ICRA 2016), 2016.
[14] Jain, et al., "Structural-RNN: Deep Learning on Spatio-Temporal
Graphs", in proc. of the 29th IEEE Conference on Computer Vision and
Pattern Recognition, 2016.
Figure 6: Samples of the Maverick planner as viewed in RViz. Left: A
suboptimal but initially feasible path with U-Turn. Right - With an obstacle
blocking the U-Turn, a K-Turn trajectory is calculated by the Maverick
planner.
278 IEEE Symposium Series on Computational Intelligence SSCI 2018

[15] Othman, et al., "Spatio-temporal Data Representation in Ontology
System for Personalized Decision Support", in proc. of the Talent
Management Symposium (TMS 2012) , Australia, July 2012.
[16] N. Kasabov, "Evolving Spiking Neural Networks for Spatio and
Spectro-Temporal Pattern Recognition", in proc. of the 2012 IEEE 6th
International Conferen ce on Intelligent Systems, pp. 27-32, 2012.
[17] Xue, et al., "PROST: Predicting Resource Usages with Spatial and
Temporal Dependencies", in proc. of the 7th ACM/SPEC International
Conference on Performance Engineering, pp 125-126, 2016.
[18] W. Li, and Y. Dang, "Human Action Recognition based on
Convolutional Neural Networks with a Convolutional Auto-Encoder", in
proc. of the 5th International Conference on Computer Sciences and
Automation Engineering (ICCSAE 2015), pp. 933-938, 2015.
[19] K. Dhoble, et al., "Online Spatio-Temporal Pattern Recognition with
Evolving Spiking Neural Networks utilising Address Event
Representation, Rank Order, and Temporal Spike Learning", in proc. of
the WCCI 2012 IEEE World Congress on Computational Intelligence,
Brisbane, Australia, 2012.
[20] M. Sànchez-Marrè et al., "Uncertainty Management, Spatial and
Temporal Reasoning, and Validation of Intelligent Environmental
Decision Support Systems", in PhD diss., International Environmental
Modelling and Software Society, 2006.
[21] J. Donahue, L. A. Hendricks, M. Rohrbach, S.Venugopalan, S.
Guadarrama, K. Saenko, and T. Darrell, “Long-term Recurrent
Convolutional Networks for Visual Recognition and Description”, in
proc. of the IEEE Conference on Computer Vision and Pattern
Recognition, pp. 2625-2634, 2015
[22] M. Sabokrou et al., "Real-time anomaly detection and localization in
crowded scenes", in proc. of IEEE Conference on Computer Vision and
Pattern Recognition Workshops, pp. 56-62, 2015.
[23] Kolonin, "Adaptive experiential learning for business intelligence
agents", in IEEE Cognitive Sciences, Genomics and Bioinformatics
(CSGB), pp. 1-5, 2016.
[24] P. Mathieu, D. Panzoli, and S. Picault, "Format-store: a multi-agent
based approach to experiential learning", in proc. of Third IEEE
International Conference on Games and Virtual Worlds for Serious
Applications (VS-GAMES), pp. 120-127, 2011.
[25] T. Somjaitaweeporn, and D. Love, "Using Intelligent Agent Technology
in a Simulation System to Enhance Experiential Learning in Operations
Management", in proc. of Sixth International Conference on Intelligent
Technologies, Phuket, pp. 270-276. 2005.
[26] F. Sahin, and J. S. Bay, "Learning from experience using a decision-
theoretic intelligent agent in multi-agent systems", in proc. of the IEEE
workshop on Soft Computing in Industrial Applications, (SMCia/01),
pp. 109-114 , 2001.
[27] D. Mladenic, “Text-learning and related intelligent agents: a survey”, in
IEEE Intelligent Systems, pp. 44–54, 1999
[28] D. Kumaran, D.Hassabis, and J. L. McClelland, “What learning systems
do intelligent agents need? complementary learning systemsz theory
updated”, in Trends in Cognitive Sciences, vol. 20, no. 7, pp. 512-534,
2016.
[29] N. B.Talley, "Imagining the Use of Intelligent Agents and Artificial
Intelligence in Academic Law Libraries", in Law Libr. J, vol .108,
p.383, 2016.
[30] C. Thaiupathump, J. Bourne, and J. Campbell, "Intelligent agents for
online learning", in Journal of Asynchronous Learning Network, vol. 3,
no. 2, p.1, 1999.
[31] C. Conati, and M.Klawe, "Socially intelligent agents in educational
games", in Socially Intelligent Agents, pp. 213-220, 2002.
[32] U.Minhas, J. Zhang, T. Tran, and R. Cohen, "Intelligent agents in
mobile vehicular ad-hoc networks: Leveraging trust modeling based on
direct experience with incentives for honesty", in proc. of 2010
IEEE/WIC/ACM International Conference Web Intelligence and
Intelligent Agent Technology (WI-IAT), vol. 2, pp. 243-247, 2010.
[33] L. Henry and S. Sankaranarayanan, "Intelligent Agent based Mobile
Learning System", in International Journal of Computer Information
Systems and Industrial Management Applications (IJCISIM), vol.2,
pp.306-319, 2010.
[34] S. A. Bagloee, M. Tavana, M. Asadi, and T. Oliver, "Autonomous
vehicles: challenges, opportunities, and future implications for
transportation policies", in Journal of Modern Transportation vol. 24,
no. 4, pp. 284-303, 2016.
[35] J. Connelly, W. S. Hong, R. B. Mahoney Jr, and D. A. Sparrow,
"Current challenges in autonomous vehicle development", in proc. of
Defense and Security Symposium, pp. 62300D-62300D, 2006.
[36] R. D'Andrea, and P. Wurman, "Future challenges of coordinating
hundreds of autonomous vehicles in distribution facilities", in proc. of
2008 IEEE International Conference on Technologies for Practical
Robot Applications, pp. 80-83, 2008.
[37] M. Campbell, M. Egerstedt, J. P. How, and R.M. Murray, "Autonomous
driving in urban environments: approaches, lessons and challenges", in
Philosophical Transactions of the Royal Society of London A:
Mathematical, Physical and Engineering Sciences, vol. 368, no. 1928,
pp. 4649-4672, 2010.
[38] C. Berger, "From autonomous vehicles to safer cars: selected challenges
for the software engineering", in proc. of International Conference on
Computer Safety, Reliability, and Security, pp. 180-189., 2012.
[39] M. Woods, R. S. Aylett, D. P. Long, M. Fox, and R. Ward, "Developing
Autonomous AI Planning and Scheduling Technologies for Remote
Planetary Exploration", in proc. of 7th ESA Workshop on Advanced
Space Technologies for Robotics and Automation (ASTRA 2002),
ESTEC, Noordwijk, The Netherlands, November 19 - 21, 2002
[40] H. Kitano, "Robocup rescue: A grand challenge for multi-agent
systems", in proc. of the Fourth IEEE International Conference on
MultiAgent Systems, pp. 5-12, 2000.
[41] K. Daniel, B. Dusza, A. Lewand owski, and C. Wietfeld, "AirShi eld: A
system-of-systems MUAV remote sensing architecture for disaster
response", in proc. of 3rd IEEE Annual Systems Conference, pp. 196-
200, 2009.
[42] R. Amant, and A. B. Wood, "Tool Use for Autonomous Agents", in
AAAI, pp. 184-189. 2005.
[43] R. Murphy, K. S. Pratt, and J. L. Burke, "Crew roles and operational
protocols for rotary-wing micro-UAVs in close urban environments", in
proc. of the 3rd ACM/IEEE international conference on Human robot
interaction, pp. 73-80, 2008.
[44] E. Hansen, T. Huntsberger, and L. Elkins, "Autonomous maritime
navigation: Developing autonomy skill sets for USVs", in proc. of
Defense and Security Symposium, pp. 62300U-62300U, 2006.
[45] E. Swanson, A. Galvao, and K. Sato. "A framework for understanding
contexts in interactive systems development." In proc. of the 7th World
Multi-Conference on Systemics, Cybernetics and Informatics, Orlando,
FL. 2003.
[46] Y. Li, and C.Zhang. "Perceiving environments for intelligent agents", in
proc. of “Pacific Rim International Conference on Artificial
Intelligence”, pp. 297-307, 2000.
[47] T. Litman, " Ready or Waiting”, in Traffic Technology International,
pp. 36-42, January 2014.
[48] D. Fagnant, and K. Kockelman. "Preparing a nation for autonomous
vehicles: opportunities, barriers and policy recommendations", in
Transportation Research Part A: Policy and Practice vol. 77, pp. 167-
181, 2015.
[49] J. E. Manley, "Unmanned surface vehicles, 15 years of development",
In OCEANS, pp. 1-4., 2008.
[50] M. Knudson, and K. Tumer, "Adaptive navigation for autonomous
robots", in Robotics and Autonomous Systems, vol. 59, no. 6, pp. 410-
420, 2011.
[51] Y. Jia, E. Shelhamer, J. Donahue, S. Karayev, J. Long, R. Girshick, S.
Guadarrama, and T. Darrell. "Caffe: Convolutional architecture for fast
feature embedding", in proc. of 22nd ACM international conference on
Multimedia, pp. 675-678, 2014.
[52] J.D. Gehrke, "Evaluating situation awareness of autonomous systems",
in Performance Evaluation and Benchmarking of Intelligent Systems,
pp. 93-111. Springer US, 2009.
[53] C. Chen, A. Seff, A. Kornhauser, and J. Xiao", Deepdriving: Learning
affordance for direct perception in autonomous driving", in proc. of
IEEE Symposium Series on Computational Intelligence SSCI 2018 279

IEEE International Conference on Computer Vision, pp. 2722-2730.
2015.
[54] D. Maturana, and S. Scherer, "Voxnet: A 3d convolutional neural
network for real-time object recognition", in proc. of 2015 IEEE/RSJ
International Conference on Intelligent Robots and Systems (IROS),pp.
922-928, 2015.
[55] Seifert, T. Bark owsky, and C. Freksa. "Region-based representation for
assistance with spatio-temporal planning in unfamiliar environments",
in Location based services and telecartography, pp. 179-191. Springer
Berlin Heidelberg, 2007.
[56] S. Mahadevan, and J. Connell, "Automatic programming of behavior-
based robots using reinforcement learning", in Artificial intelligence,
vol. 55, no. 2-3, pp. 311-365, 1992.
[57] B. Sellner, F.W. Heger, L.M. Hiatt, R. Simmons, and S. Singh,
"Coordinated multiagent teams and sliding autonomy for large-scale
assembly", in IEEE - Special Issue On Multi-Robot.
[58] E. Ohn-Bar, and M. Trivedi. "Are all objects equal? Deep spatio-
temporal importance prediction in driving videos." Pattern Recognition
64 (2017): 425-436.
[59] Y. Jang, Y. Song, Y. Yu, Y. Kim, and G. Kim. "TGIF-QA: Toward
spatio-temporal reasoning in visual question answering." In IEEE
Conference on Computer Vision and Pattern Recognition (CVPR 2017).
Honolulu, Hawaii, pp. 2680-8. 2017.
[60] F. Lattari, M. Ciccone, M. Matteucci, J. Masci, and F. Visin.
"ReConvNet: Video Object Segmentation with Spatio-Temporal
Features Modulation." arXiv preprint arXiv:1806.05510 (2018).
[61] J. Xu, S. Wang, S. Su, SAP Kumar, and C. Wu. "Latent Interest and
Topic Mining on User-Item Bipartite Networks." In 2016 IEEE
International Conference on Services Computing (SCC), pp. 778-781.
IEEE, 2016.
[62] B. Xu, and S.A. Kumar. "Big Data Analytics Frame work for System
Health Monitoring." In Big Data (BigData Congress), 2015 IEEE
International Congress on, pp. 401-408. IEEE, 2015.
[63] B. Xu, and S. Kumar. "A Text Mining Classification Framework and its
Experiments Using Aviation Datasets." (2015).
[64] D. Eastman and S. A. P. Kumar, "A Simulation Study to Detect Attacks
on Internet of Things," 2017 IEEE 15th Intl Conf on Dependable,
Autonomic and Secure Computing, 15th Intl Conf on Pervasive
Intelligence and Computing, 3rd Intl Conf on Big Data Intelligence and
Computing and Cyber Science and Technology
Congress(DASC/PiCom/DataCom/CyberSciTech), Orlando, FL, 2017,
pp. 645-650.
[65] S. A. Kumar, T. Vealey, and H. Srivastava. "Security in internet of
things: Challenges, solutions and future directions." In System Sciences
(HICSS), 2016 49th Hawaii International Conference on, pp. 5772-
5781. IEEE, 2016.
[66] C. Chrane, and S. Kumar. "An Examination of Tor Technology Based
Anonymous Internet." In Proceedings of the 15th Informing Science
Institute (InSite) International Conference, Tampa, FL. 2015.
[67] J. Liu, S. Wang, A. Zhou, S. Kumar, F.Yang, and R. Buyya. "Using
proactive fault-tolerance approach to enhance cloud service reliability."
IEEE Transactions on Cloud Computing (2016).
[68] X. Wei, S. Wang, A. Zhou, J. Xu, S.Su, S. Kumar, and F. Yang. "MVR:
An Architecture for Computation Offloading in Mobile Edge
Computing." In 2017 IEEE International Conference on Edge
Computing (EDGE), pp. 232-235. IEEE, 2017.
[69] S. Kumar, and B. Xu. "Vulnerability Assessment for Security in
Aviation Cyber-Physical Systems." In Cyber Security and Cloud
Computing (CSCloud), 2017 IEEE 4th International Conference on, pp.
145-150. IEEE, 2017.
[70] S. Kumar, B. Bhargava, R. Macêdo, and G. Mani. "Securing IoT-Based
Cyber-Physical Human Systems against Collaborative Attacks." In
Internet of Things (ICIOT), 2017 IEEE International Congress on, pp. 9-
16. IEEE, 2017.
[71] S. Revels, S. Kumar, and O. Ben-Assuli. "Predicting obesity rate and
obesity-related healthcare costs using data analytics." Health Policy and
Technology 6, no. 2 (2017): 198-207.
280 IEEE Symposium Series on Computational Intelligence SSCI 2018