=Paper=
{{Paper
|id=Vol-3261/paper9
|storemode=property
|title=Unsupervised continual learning from synthetic data generated with agent-based modeling and simulation: a preliminary experimentation
|pdfUrl=https://ceur-ws.org/Vol-3261/paper9.pdf
|volume=Vol-3261
|authors=Gianfranco Lombardo,Mattia Pellegrino,Agostino Poggi
|dblpUrl=https://dblp.org/rec/conf/woa/LombardoPP22
}}
==Unsupervised continual learning from synthetic data generated with agent-based modeling and simulation: a preliminary experimentation==
https://ceur-ws.org/Vol-3261/paper9.pdf
Unsupervised Continual Learning From Synthetic
Data Generated with Agent-Based Modeling and
Simulation: A preliminary experimentation
Gianfranco Lombardo1,*,† , Mattia Pellegrino1,† and Agostino Poggi1,†
1
University of Parma, Parma, 43125, Italy
Abstract
Continual Learning enables to learn a variable number of tasks sequentially without forgetting knowledge
obtained from the past. Catastrophic forgetting usually occurs in neural networks for their inability to
learn different tasks in sequence since the performance on the previous tasks drops down in a significant
way. One way to solve this problem is providing a subset of the previous examples to the model while
learning a new task. In this paper we evaluate the continual learning performance of an unsupervised
model for anomaly detection by generating synthetic data using an Agent-based modeling and simulation
technique. We simulated the movement of different types of individuals in a building and evaluate their
trajectories depending on their role. We collected training and test sets based on their trajectories. We
included, in the test set, negative examples that contain wrong trajectories. We applied a replay-based
continual learning to teach the model how to distinguish anomaly trajectories depending on the users’
roles. The results show that using ABMS synthetic data it is enough a small percentage of synthetic data
replay to mitigate the Catastrophic Forgetting and to achieve a satisfactory accuracy on the final binary
classification (anomalous / non-anomalous).
Keywords
Agent-based modeling, Synthetic data, Data generation, Continual Learning, Replay methods,
1. Introduction
Deep Learning has been growing in scale, breadth of applications, and the amount of required
data, the so-called “data-hungry" deep neural networks. This is an important problem to deal
with, especially when the task to be learned presents dynamic aspects in terms of environment
and classes of tasks where few data are available during the training phase. Moreover, for
some context, new tasks become available only during the time and for this reason, the models
should be smoothly updated. When learning new tasks from data in sequence, Neural Networks
usually face the so-called Catastrophic Forgetting: performance on the previous tasks drops
significantly. This is a serious disadvantage that prevents many intelligent applications to
real-life problems when not all the context is known beforehand. Continual Learning (CL) aims
WOA 2022: 23rd Workshop From Objects to Agents, September 1–2, Genova, Italy
*
Corresponding author.
†
These authors contributed equally.
$ gianfranco.lombardo@unipr.it (G. Lombardo); mattia.pellegrino@unipr.it (M. Pellegrino);
agostino.poggi@unipr.it (A. Poggi)
0000-0002-0877-7063 (G. Lombardo); 0000-0001-7116-9338 (M. Pellegrino); 0000-0002-9421-8566 (A. Poggi)
© 2022 Copyright for this paper by its authors. Use permitted under Creative Commons License Attribution 4.0 International (CC BY 4.0).
CEUR
Workshop
Proceedings
http://ceur-ws.org
ISSN 1613-0073
CEUR Workshop Proceedings (CEUR-WS.org)
to introduce novel methodologies to enable the update of a model in terms of new tasks and
data distributions, unseen in the past, without losing the already acquired knowledge to solve
the previous tasks accurately [1].
On the other hand, another possibility to have a higher quantity of data or enough data for
new tasks is exploiting synthetic data generation techniques [2]. More recently, it has also been
proven that using synthetic data during training can reduce Catastrophic Forgetting when used
for a Continual Learning framework [3].
In several contexts, Agent-based Modeling and Simulation (ABMS) techniques can repre-
sent an effective way to generate synthetic data. In this paper, we propose the preliminary
experimentation performed using Continual Learning with synthetic data generated using
ABMS. In particular, we investigated an unsupervised reconstruction task of high-resolution
trajectories that aims to recognize the anomaly trajectories followed by the users inside a
building depending on their job role. Users are divided in three classes: customers, workers
and maintenance operators. Trajectories of each users have been generated exploiting the
Boid model [4] and Agent-based Modeling and simulation techniques. An unsupervised Deep
Recurrent Neural network is trained to recognize the anomalies in the trajectories of each class
of users by exploiting a Replay-based Continual Learning technique.
2. Literature review
To correctly generate synthetic data to exploit the characteristics of continuous (lifelong)
learning, we first had to choose a model suitable for our use case. We considered the main
modeling and simulation techniques used for crowd and pedestrian modeling within the current
literature, at different scales of complexity. A large number of models are available to implement
simulations regarding pedestrians inside or outside buildings and their relative behavior. These
models can be classified into two main classes:
• macroscopic models: focus on the whole system as a single entity (regression models,
route choice models)
• microscopic models: study the behavior, decision, and attributes of the individuals and
their interaction among themselves (rule models, cellular automata).
Regression models analyzed crowd relations in order to predict pedestrian flow behavior
under specific circumstances. This method is strong infrastructure bound [5]. In [6] Nassereddine
et al. used a non-probabilistic regression approach to model the behavior of right-turning drivers:
when drivers perceive the possibility of a pedestrian reaching a critical conflict point at the
same time as them, they will modify their behavior, even if not coming to a complete stop.
In route choice models pedestrians choose their destinations in order to maximize the utility
of their path. The utility defines a set of properties: travel time, points visited, etc. In [7]
Alivand et al. model a scenic route between two locations identifying several variables of
the surrounding environment as significant contributors to route scenery, they considered a
scenario regarding tourist and recreation travelers. They used, more specifically, a Path Size
Logit (PSL) model, to identify the relevant attributes and their relative importance.
Cellular Automata divide the space into a uniform grid. Each agent owns a particular grid
position, a single grid position is called "cell". Each agent can move between cells depending
on the modeling system. Cellular Automata use discrete time steps to measure the time. The
variables at each cell are updated simultaneously based on the values of the variables in their
neighborhood at the previous time step [8]. In [9] Zhao et al. used a 2D random cellular automata
to model the occupant evacuation considering the influence of human psychology and behavior.
Rule-based models have been used to simulate flocks of animals or crowds of people. A well-
known model that is able to simulate life-like complex behavior is the boid model [10]. Each
agent is an independent entity that navigates according to its perception and the environment
where is placed, the physics, and its behavior. The aggregate motion of flocks is created by such
distributed model. To simulate a generic flocking behavior, the basic rules are just three: they
describe how an entity has to move, taking into account its position, its velocity, and how near
to another entity is:
• separation: steer to avoid crowding local flock mates
• alignment: steer towards the average heading of the local flock mates
• cohesion: steer towards the average position of the local flock mates
In order to use these three rules, it’s important to take into consideration not only the position of
an entity but even its neighborhood. So it must be considered that managing the neighborhood
can also be very complex. Each entity moves and therefore must be considered along with
all the other entities within a certain radius. The computation complexity is 𝑂(𝑛2 ), where
𝑛 is the number of entities in the neighborhood. To simplify and speed up this operation, a
neighborhood can be modeled in a simpler way. Each agent can move in a limited range, so they
can move a lot over time. The basic idea is to divide the space into a set of bins. The steps are:
• Create a circle or sphere (2D or 3D)
• See if the bin falls into the sphere
• Follow the rules for cohesion, separation, and alignment
In [11] Shinoda et al. designed and developed a pedestrian dynamics simulator to perform
navigation system studies. Moreover, they performed simulation experiments with a novel
wearable navigation device that can guide a pedestrian through an evacuation process.
In the light of these analyses, we decided to use the boid approach to model our use case.
Moreover, we are interested to simulate a complex scenario at a microscopic scale. The agents,
involved in the simulation process, have to reach predetermined points and reach gradually the
end of their path. For these reasons, we decided to adopt a rule-based model (boid) because it is
well-suit for our needs.
After choosing the most suitable simulation method, our attention shifted to the trajectory
data generation, produced by the system, and the exploitation of these data for anomaly detection
tasks and continual (lifelong) learning.
Detecting an anomaly behavior is an important problem in various situations. Finding an
anomalous behavior may indicate important objects and events in a wide variety of domains. In
literature, there is a huge variety of algorithms that can detect an anomalous behavior during
a predetermined trajectory [12]. In [13] Laxhammar et al. proposed a Sequential Hausdorff
Nearest-Neighbor Conformal Anomaly Detector (SHNN-CAD) for online learning and sequential
anomaly detection in trajectories. However, in this work, we adopted an encoder-decoder
architecture based on Long-Short Term Memory (LSTM) to perform the anomaly detection
task. We chose LSTM among other models for its ability to deal with temporal sequence and
maintain the temporal dependencies [14].
Continual learning provides the ability to learn continuously and enrich its ability when data
and tasks become available over time [1]. Continual Learning updates its intelligent model
taking into account data and attributes that it didn’t see in the past training steps, without
losing any information about the past (Catastrophic Forgetting) [15].
To minimize the forgetting phenomenon, in a real use case, it might be useful to use synthetic
data along with the real ones. In [16] Ma et al. trained a multi-agent future trajectories
predictor given a stream of datasets collected at different scenarios. In particular, they use a
graph-neural-network-based conditional generative memory system to mitigate catastrophic
forgetting. However, in scientific literature there is currently a lack of unsupervised models
trained with a CL paradigm. Indeed, most of the state of the art techniques are related to
supervised tasks.It is possible to distinguish three categories of CL techniques:
• Replay-based methods: When learning a new task, the model is exposed with a variable
percentage to examples from the previous learned tasks [17, 18]
• Regularization-based methods: Continual learning is performed by acting on the parame-
ters of the model by discouraging the updating of neural network’s layers that are deemed
relevant for the single tasks. Unfortunately, these methods does not scale well when
increasing the number of tasks [19].
• Architectural methods: Developing ad-hoc solutions where different parts of the model
take care of each task exclusively [20, 21]
3. Use case
As use-case to investigate continual learning from ABMS we choose an anomaly detection task
related to people’s trajectories inside a building as a service offered by an intelligent Location-
based service. We hypothesized an indoor localization infrastructure that locates individuals
using different tracking technologies with different sample and errors rate.
We focused on the development of an intelligent service which exploits Location-based
service data. Indeed, Locate and analyze indoor users’ paths may introduce novel intelligent
services that can help management to make predictions, measure performance indicators
and ensure safety conditions. Examples of services could be: Measuring the inefficiency of
spatial organization, occupation of spaces prediction, users waiting times estimation and finally
anomaly detection.
We focus on indoor users’ paths to classify if they are allowed to visit specific environments
depending on their role or to understand if any change in the building organization induces
anomaly trajectories for a set of users that have the same role in the organization. Different
classes of users can emerge over time and are required to be learned within their normal
behaviors. Depending on the target organization the roles can be different: for example, in a
hospital scenario could be medical staff, patients, and maintenance staff; In a retail environment
could be customers, salesmen, and warehouse workers. Modeling and simulating the users’
behaviors offer mainly two advantages:
• It is possible to generate synthetic trajectories when a new category of users should be
tracked and recognized and there is not real-data available to train the anomaly detection
model on this new one
• It is possible to evaluate how the performance of the intelligent service changes depending
on the accuracy and sample rate that is used to locate and track the class of users.
This anomaly detection task can be performed using both supervised or unsupervised ML
techniques. However, when users’ tracking is performed with a high-resolution as a sequence
of visited locations in the environment [(𝑥0 , 𝑦0 )...(𝑥𝑇 , 𝑦𝑇 )] several benefits come up with the
adoption of an unsupervised paradigm like auto encoder-based trajectory reconstruction:
• No need of a priori knowledge of the environment topology nor supervision, since allowed
trajectories can be learnt by observing a set of trusted real users or with agents that
perform trustable trajectories;
• Easy adoption in a new organization;
• Low cost to update the model if something changes in the environment topology
Figure 1 shows a map segmentation of a real building we used for the experiments. The
environment is divided according to the main operations that are daily performed. In our use-
case the building is a real-existing hospital ward where can interact three different categories of
users: patients that need for treatments; medical staff and finally the maintenance operators.
Patients enter the ward (yellow area) by the waiting room and proceed through the back-office
and reception area (green area). Patients can also visit the clinics areas (orange) and the areas
where treatments are provided (pink areas). They cannot access the maintenance area (violet
zone) and the private spaces of clinicians and the archive (Red and purple areas).
On the other hand, the medical staff can access all the environments except the maintenance
area and can enter the ward from multiple gates. Finally, the maintenance staff can enter from
multiple gates, they can access the private area for maintenance but we supposed they cannot
access the archive and the private medical offices.
4. ABMS
To realize our ABMS model we exploited the properties of ActoDeS. ActoDeS is a software
framework that allows the development of distributed and concurrent systems [22]. ActoDeS
is Java-based and inherits some functionalities and some implementation solutions used in
JADE [23, 24, 25]. Multiple and various application has been developed with the aid of ActoDeS
functionalities, especially for agent-based modeling and simulation [26, 27, 28], data analysis
[29, 30, 31, 32, 33], and more recently for epidemiological simulation, [34], [35]. This framework
is based on the use of concurrent objects (from here named actors) whose main characteristics
derive from the actor model [36]. In ActoDeS, an actor is created by another actor; after
its creation, the actor can interact with other actors through the exchange of asynchronous
messages and, as a consequence, can change its behavior several times; once its work is finished,
the actor kills itself. ActoDeS can also support millions of actors and simulate their behavior.
Indeed, it has a special structure in which agents can be divided into different pools and managed
using multiple threads and resources.
Figure 1: The real-hospital ward map used for the experimental part. The map is color-based segmented
according to the daily activities. The colored dots are the gravitational attraction point we used for the
Boid simulation using ActoDeS.
In order to develop our use case, we create a virtual environment, where the agents can move
and interact among themselves. Each agent represents an individual (actor), that can have a
specific role and a specific path. Moreover, it could be subject to limitations and can’t access
in some specific areas of the virtual environment. We simulate the whole trajectory set by
exploiting the Boid model. We can have different types of boid depending on how many agents’
categories we want to implement. To create the virtual environment and limit the 2D space in
which placing boids we defined as "wall" type actors. These particular types of actors have their
own dimension in space and have the task of limiting the movements of the other "pedestrian"
actors in a well-defined space.
The agents can reach their location randomly follow other agents or choose between some
alternative paths. Each agent is assigned a starting point, a goal point, and a list of waypoints
to cross, always according to their category. To generate the most random simulations possible,
we planned to decentralize the starting point of every single agent by a certain offset randomly
generated. Each trajectory has different lengths.
With this particular type of approach, we can generate different types of boid and different
types of professional figures who, for example, move independently in the workplace. Thanks
to the ABMS architecture, is possible to generate synthetic data for the training and test set
of the application we want to create. Moreover, when we create the test set, we also include
negative examples which can contain trajectories for paths not allowed for certain types of boid.
Finally, by setting the sampling rate it is possible to simulate a data gathering using different
indoor localization techniques.
5. Continual Learning framework
The machine learning model has no prior knowledge about the environment and the allowed
trajectories for each category but it learns the allowed ones in the environment for each category
as an anomaly detection task over spatial location sequences. The only prior knowledge
exploited by the model is the user’s role that in a real context would be retrieved according to
the technology used to track the user.
The machine learning model learns to reconstruct the high-resolution trajectory of each user
using and encoder-decoder architecture aiming to reduce the Root Mean Squared Error (RMSE)
among the input sequence and the reconstructed one. This is a common methodology used to
detect anomalies in sequences and time-series because the model fails to reconstruct sequence
that rely on distributions that have not be seen during the training phase with a higher RMSE.
We also provided the user role as input of the model together with its trajectory. We exploited an
encoder-decoder architecture based on Long-Short Term Memory (LSTM) deep neural networks.
LSTM network have been selected because of their ability of learning patterns on sequence
data by considering and preserving temporal dependencies [37]. The encoder network process
the trajectory as a sequence of 150 visited locations (x,y) along with a categorical variable that
identifies the user category and learns an internal compressed representation that is then used
by the decoder network to reconstruct the trajectory. Figure 2 shows the main points of the
architecture.
A trajectory is considered anomaly when the Root Mean Squared Error (RMSE) of the two
sequences is above a certain threshold that is estimated using a validation set. According to
the RMSE, the model classifies a trajectory for a user as not allowed when it results to be
anomaly and allowed otherwise. One challenge when dealing with such task with unsupervised
techniques is teaching the model that the same trajectory can be for example normal and
allowed if the user belongs to the medical staff but not if he belongs to another category without
supervising the training. That is the reason why we provide the categorical variable that
indicates the user’s category in input to the encoder along with its trajectory.
With this hypothesis we experimented this use-case adopting a Continual Lifelong learning
paradigm. The model is firstly trained to only reconstruct normal trajectories of doctors which
represents the first task. After that the neural network is trained to deal with a second task:
reconstruct normal trajectories of the maintenance staff and finally the patients’ trajectories. The
model is fine-tuned using the validation-set but results have been computed using a balanced
test set with the same numbers of normal and anomaly trajectories.
We analyzed how the performances of classifying anomalies change by incrementally learning
all the tasks separately. The main goal of this experiment is avoid and reduce the so-called
Catastrophic Forgetting of the model when learning new tasks. In our case, this goal becomes
more challenging since all tasks are learnt in an unsupervised way. We exploited a Replay-based
method by analyzing the performances when changing the percentage of replay samples (𝑅% ).
The method is composed by the following steps:
1. We first trained the model using only the 100% of doctors’ trajectories available (Traditional
training) and performing the fine-tune to select the reconstruction error threshold using
the validation set.
Reconstructed trajectory
x x1 x2 x3 ......... xs
y y1 y2 y3 ......... ys
Output
LSTM1 LSTM2 LSTMK
Decoder network with k units
Encoded features
LSTM1 LSTM2 LSTMn
Encoder network with n units
User trajectory
x x1 x2 x3 ......... xs
y y1 y2 y3 ......... ys
Role 0 0 0 .......... 0
Figure 2: LSTM Encoder-Decoder architecture used to reconstruct each trajectory. The image also
shows the structure we adopted for the input and the desired output.
2. Starting from the previous model, the neural network is trained to learn the second
task (reconstruct normal trajectories of maintenance staff), by using the 100 % available
samples for this new task and an amount of examples of the previous task equal to 𝑅%
3. Finally, the third task (reconstruct normal trajectories of patients) is learnt in the same
way as the second one: 100 % available samples for the new task and some additional
examples for the previous tasks to avoid the Catastrophic Forgetting of the network.
Respectively, the 𝑅% of the available doctors’ normal trajectories and of the available
maintenance normal trajectories.
For each configuration we performed 50 runs and report the average performances. Since
the model incrementally learns to reconstruct the trajectories of a class of users we analyzed
the Precision, Recall and F-1 score achieved by the model when classifying anomalies for each
category. Figure 3 shows the result.
6. Results
If trained with a replay-percentage 𝑅% equal to zero, Catastrophic Forgetting occurs and only
the third task is learnt. Introducing a 𝑅% = 1%, performances improve with a final accuracy in
the binary task (normal/anomaly) equal to 62.7%. At the level of the single user category the
one that suffers mostly a bad performance is the one related to the maintenance staff.
We then tested with 𝑅% = 10, 30, 50%. It is interesting that by only replaying the 10% of
examples of the first two tasks the final binary accuracy reaches on average the 84.5% and
the slope of improvements at the single categories level becomes less pronounced. Finally, we
tested using 𝑅% = 100 to compare with a traditional training that achieves a binary accuracy
of 91.08 when learning with the a priori knowledge of all the classes of users. The incremental
one reaches on average an accuracy equal to 90.8%. It is clear that there is not any important
difference by training incrementally over the single categories if the replay involves all the
available examples. However, in terms of F-1 score the patient category is slightly learnt better
with a traditional learning paradigm.
Figure 3: F-measure, Precision and Recall for each task when using a Replay methods for the Continual
Learning scenario. Tasks are learnt in the order: Doctors,maintenance and finally patients. Results are
reported as an average of different runs with respect of the replay percentage used.
7. Conclusions
In this paper we presented the first experiments we performed with replay-based continual
learning using ABMS to generate synthetic data. Moreover, we propose a unsupervised CL
approach for anomaly detection of trajectories depending on the users’ role. The results are
promising and by only using a small percentage of replay it is possible to achieve an overall
good accuracy in the final classification task. In our future works we will investigate other
CL approaches to understand if they also benefit from ABMS synthetic data. Moreover, we
will further analyze a more complex scenario that can involve more than three tasks (users’
categories to be recognized as anomaly or not).
References
[1] G. I. Parisi, R. Kemker, J. L. Part, C. Kanan, S. Wermter, Continual lifelong learning with
neural networks: A review, Neural Networks 113 (2019) 54–71.
[2] S. I. Nikolenko, Synthetic data for deep learning, volume 174, Springer, 2021.
[3] W. Masarczyk, I. Tautkute, Reducing catastrophic forgetting with learning on synthetic
data, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern
Recognition Workshops, 2020, pp. 252–253.
[4] G. Angiani, P. Fornacciari, G. Lombardo, A. Poggi, M. Tomaiuolo, Actors based agent
modelling and simulation, in: International Conference on Practical Applications of Agents
and Multi-Agent Systems, Springer, 2018, pp. 443–455.
[5] J. S. Milazzo, II, N. M. Rouphail, J. E. Hummer, D. P. Allen, Effect of pedestrians on capacity
of signalized intersections, Transp. Res. Rec. 1646 (1998) 37–46.
[6] H. Nassereddine, K. R. Santiago-Chaparro, D. A. Noyce, Modeling vehicle–pedestrian
interactions using a nonprobabilistic regression approach, Transp. Res. Rec. 2675 (2021)
356–364.
[7] M. Alivand, H. Hochmair, S. Srinivasan, Analyzing how travelers choose scenic
routes using route choice models, Computers, Environment and Urban Systems 50
(2015) 41–52. URL: https://doi.org/10.1016%2Fj.compenvurbsys.2014.10.004. doi:10.1016/
j.compenvurbsys.2014.10.004.
[8] N. Pelechano, A. Malkawi, Evacuation simulation models: Challenges in modeling high
rise building evacuation with cellular automata approaches, Automation in Construction
17 (2008) 377–385. URL: https://doi.org/10.1016%2Fj.autcon.2007.06.005. doi:10.1016/j.
autcon.2007.06.005.
[9] D. L. Zhao, J. Li, Y. Zhu, L. Zou, The application of a two-dimensional cellular automata
random model to the performance-based design of building exit, Build. Environ. 43 (2008)
518–522.
[10] C. W. Reynolds, Flocks, herds and schools: A distributed behavioral model, in: Pro-
ceedings of the 14th annual conference on Computer graphics and interactive tech-
niques - SIGGRAPH '87, ACM Press, 1987. URL: https://doi.org/10.1145%2F37401.37406.
doi:10.1145/37401.37406.
[11] K. Shinoda, I. Noda, E. Oyama, A pedestrian dynamics simulator for wearable navigation
(????).
[12] S. A. Ahmed, D. P. Dogra, S. Kar, P. P. Roy, Trajectory-based surveillance analysis: A survey,
IEEE transactions on circuits and systems for video technology 29 (2018) 1985–1997.
[13] R. Laxhammar, G. Falkman, Online learning and sequential anomaly detection in trajecto-
ries, IEEE transactions on pattern analysis and machine intelligence 36 (2013) 1158–1173.
[14] Y. Yu, X. Si, C. Hu, J. Zhang, A review of recurrent neural networks: LSTM cells and
network architectures, Neural Comput. 31 (2019) 1235–1270.
[15] R. Kemker, M. McClure, A. Abitino, T. Hayes, C. Kanan, Measuring catastrophic forgetting
in neural networks, in: Proceedings of the AAAI Conference on Artificial Intelligence,
volume 32, 2018.
[16] H. Ma, Y. Sun, J. Li, M. Tomizuka, C. Choi, Continual multi-agent interaction behavior
prediction with conditional generative memory, IEEE Robotics and Automation Letters 6
(2021) 8410–8417.
[17] P. Buzzega, M. Boschini, A. Porrello, D. Abati, S. Calderara, Dark experience for general
continual learning: a strong, simple baseline, Advances in neural information processing
systems 33 (2020) 15920–15930.
[18] D. Rolnick, A. Ahuja, J. Schwarz, T. Lillicrap, G. Wayne, Experience replay for continual
learning, Advances in Neural Information Processing Systems 32 (2019).
[19] F. Zenke, B. Poole, S. Ganguli, Continual learning through synaptic intelligence, in:
International Conference on Machine Learning, PMLR, 2017, pp. 3987–3995.
[20] J. Schwarz, W. Czarnecki, J. Luketina, A. Grabska-Barwinska, Y. W. Teh, R. Pascanu,
R. Hadsell, Progress & compress: A scalable framework for continual learning, in:
International Conference on Machine Learning, PMLR, 2018, pp. 4528–4537.
[21] G. Lombardo, A. Poggi, M. Tomaiuolo, Continual representation learning for node classifi-
cation in power-law graphs, Future Generation Computer Systems 128 (2022) 420–428.
[22] F. Bergenti, E. Franchi, A. Poggi, Agent-based social networks for enterprise collaboration,
in: 2011 IEEE 20th International Workshops on Enabling Technologies: Infrastructure for
Collaborative Enterprises, IEEE, 2011, pp. 25–28.
[23] A. Poggi, M. Tomaiuolo, P. Turci, Extending jade for agent grid applications, in: 13th
IEEE International Workshops on Enabling Technologies: Infrastructure for Collaborative
Enterprises, IEEE, 2004, pp. 352–357.
[24] A. Negri, A. Poggi, M. Tomaiuolo, P. Turci, Dynamic grid tasks composition and distribution
through agents, Concurrency and Computation: Practice and Experience 18 (2006) 875–885.
[25] F. Bergenti, G. Caire, S. Monica, A. Poggi, The first twenty years of agent-based software
development with jade (2020).
[26] A. Poggi, Agent based modeling and simulation with actomos., in: WOA, 2015, pp. 91–96.
[27] F. Bergenti, A. Poggi, M. Tomaiuolo, An actor based software framework for scalable
applications, in: International Conference on Internet and Distributed Computing Systems,
Springer, 2014, pp. 26–35.
[28] M. Pellegrino, G. Lombardo, M. Mordonini, M. Tomaiuolo, S. Cagnoni, A. Poggi, Actodemic:
A distributed framework for fine-grained spreading modeling and simulation in large scale
scenarios., in: WOA, 2021, pp. 194–209.
[29] G. Lombardo, P. Fornacciari, M. Mordonini, M. Tomaiuolo, A. Poggi, A multi-agent
architecture for data analysis, Future Internet 11 (2019) 49.
[30] E. Franchi, A. Poggi, M. Tomaiuolo, Blogracy: A peer-to-peer social network, in: Censor-
ship, Surveillance, and Privacy: Concepts, Methodologies, Tools, and Applications, IGI
global, 2019, pp. 675–696.
[31] E. Franchi, A. Poggi, M. Tomaiuolo, Social media for online collaboration in firms and orga-
nizations, in: Information Diffusion Management and Knowledge Sharing: Breakthroughs
in Research and Practice, IGI Global, 2020, pp. 473–489.
[32] G. Lombardo, A. Poggi, Actornode2vec: An actor-based solution for node embedding over
large networks, Intelligenza Artificiale 14 (2020) 77–88.
[33] G. Lombardo, A. Poggi, A scalable and distributed actor-based version of the node2vec
algorithm., in: WOA, 2019, pp. 134–141.
[34] M. Pellegrino, G. Lombardo, S. Cagnoni, A. Poggi, High-performance computing and abms
for high-resolution covid-19 spreading simulation, Future Internet 14 (2022) 83.
[35] G. Lombardo, M. Pellegrino, M. Tomaiuolo, S. Cagnoni, M. Mordonini, M. Giacobini,
A. Poggi, Fine-grained agent-based modeling to predict covid-19 spreading and effect of
policies in large-scale scenarios, IEEE Journal of Biomedical and Health Informatics 26
(2022) 2052–2062.
[36] G. A. Agha, Actors: A model of concurrent computation in distributed systems., Technical
Report, Massachusetts Inst of Tech Cambridge Artificial Intelligence Lab, 1985.
[37] Y. Yu, X. Si, C. Hu, J. Zhang, A review of recurrent neural networks: Lstm cells and network
architectures, Neural computation 31 (2019) 1235–1270.