Collective adaptive systems (CASs) are challenging from the engineering perspective. Different techniques aim at taming these systems, either using declarative or black-box approaches (e.g. Machine Learning, Evolutionary Algorithms, etc.).Among the many declarative approaches, Aggregate Computing is a novel technique by which developers can express collective system behaviours from a global perspective, using a compositional and functional programming technique. Over the years, Aggregate Computing has been applied in different scenarios, ranging from smart cities to a crowd of augmented people. Despite its promising capabilities, it is sometimes challenging to describe aggregate behaviours, so we aim at merging Aggregate Computing with black-box techniques to simplify the aggregate program synthesis
Doctoral Symposium ACSOS 2021: Research directions for Aggregate Computing with Machine Learning
1. Research directions for Aggregate Computing with
Machine Learning
Gianluca Aguzzi1
Supervisor : Mirko Viroli 1
ACSOS Mentor : Christopher Landauer
1
Alma Mater Studiorum – Università di Bologna, Cesena, Italy
October 1, 2021
Talk @ ACSOS 2021, Doctoral Symposium
2nd IEEE International Conference on Autonomic Computing and Self-Organizing Systems
2. Background
Collective (Self-)Adaptive Systems (CSAS) [D’A+19]
Distributed and interconnected systems composed of multiple agents that can perform complex tasks ex-
hibiting robust collective behaviours while achieving system-wide and agent-specific goals.
Background G. Aguzzi 2 / 20
3. Aggregate Computing [BPV15]
A top-down global-to-local approach to express
collective behaviour
Rooted on field-calculus [Aud+18]
Collective behaviour does not depend on system scale
Used in various scenarios ranging from smart cities to
crowd engineering [Cas+19]
Problem i
Building block design is complex
Background G. Aguzzi 3 / 20
4. Machine Learning
Enhance agents with some learning capabilities
Learning improve adaptability, helping agent to act in uncertain environments
Supervised Learning [D’A+19], Reinforcement Learning [HW98; NNN20; MLF07;
HKT19] and Evolutionary Computing [JMK17; PL05] are typically used in CSAS
Problem i
Solutions are application-specific
Background G. Aguzzi 4 / 20
5. Problem statement
Extend Aggregate Computing paradigm to include the Machine Learning capability to improve adapt-
ability and simplify the definition and refinement of collective behaviour.
Problem statement G. Aguzzi 5 / 20
6. Motivation
¢ Aggregate Computing is scale independent by construction
¢ Hybrid collective program description
¢ Try to improve current state-of-the-art Machine Learning applying in CSAS
Problem statement G. Aguzzi 6 / 20
7. Research Questions
? What kind of Machine Learning approach is useful in combination with Aggregate
Computing
? At what level of abstraction can Machine Learning be useful for Aggregate Computing
? What does Aggregate Computing have in common with Machine Learning, applied to
Collective Self-Adaptive System
Problem statement G. Aguzzi 7 / 20
8. Early results
Setting
Focus on simple but well-known
problems in Aggregate Computing
Learning exploited to guide
building-block improvements
Verifying what kind of approach is
well-suited for Aggregate Computing
Constraints
Learning problem framed as
Homogenous Team Learning [PL05]
Learning performed off-line [PMV13]
Early results G. Aguzzi 8 / 20
9. Computational model [HLM15]
Ensemble of nodes with an
identifier
Each node has a local-view (i.e.
neighbours relationship)
Interaction happens with
message passing (executed
continously).
Early results G. Aguzzi 9 / 20
10. Computational model
Round steps
1 Context creation
2 Program evaluation producing an
export
3 Export sharing to neighbourhood
Context
Early results G. Aguzzi 10 / 20
11. Gradient/Hop count example [Aud+17]
Definition
A program that produce a computational field where each node contains the distance from a
source zone.
S
10
15
20
15
0
∞
∞
∞
∞
S
10
15
20
15
0
∞
∞
∞
∞
(∞, ∞, ∞, 0)
out = 0 + 10
S
10
15
20
15
0
10
25
25
30
Early results G. Aguzzi 11 / 20
13. Aggregate Computing + Reinforcement Learning
Most suitable match
Typically used in CSAS
Designed to maximise long term rewards
Quite easy to express aggregate problems as reinforcement
learning problems
Early results G. Aguzzi 13 / 20
14. Aggregate Computing + Reinforcement Learning: Hop count
Independent Q-Learning
Actions Increase / NoOp
State Temporal Windows of Speed
Reward 0 / Ñ -1
output
Increase
1
NoOp
0
State
[0, 1, 1, 1]
0
-1
Early results G. Aguzzi 14 / 20
15. Conclusion
Endow Aggregate Computing with intelligent behaviour
Reinforcement Learning matches with Aggregate Computing
A path towards robust, adaptive and smarter collective behaviour definition
Conclusion G. Aguzzi 15 / 20
16. References I
[D’A+19] Mirko D’Angelo et al. “On learning in collective self-adaptive systems: State of practice
and a 3D framework”. ICSE Workshop on Software Engineering for Adaptive and
Self-Managing Systems 2019-May (2019), pp. 13–24. ISSN: 21567891. DOI:
10.1109/SEAMS.2019.00012.
[Aud+18] Giorgio Audrito et al. “Space-Time Universality of Field Calculus”. Ed. by Giovanna
Di Marzo Serugendo and Michele Loreti. Vol. 10852. Lecture Notes in Computer Science.
Springer, 2018, pp. 1–20. DOI: 10.1007/978-3-319-92408-3_1.
[Cas+19] Roberto Casadei et al. “Self-organising Coordination Regions: A Pattern for Edge
Computing”. Ed. by Hanne Riis Nielson and Emilio Tuosto. Vol. 11533. Lecture Notes in
Computer Science. Springer, 2019, pp. 182–199. DOI: 10.1007/978-3-030-22397-7_11.
[BPV15] Jacob Beal, Danilo Pianini, and Mirko Viroli. “Aggregate Programming for the Internet of
Things”. Computer 48.9 (2015), pp. 22–30. DOI: 10.1109/MC.2015.261.
[HW98] Junling Hu and Michael P. Wellman. “Multiagent Reinforcement Learning: Theoretical
Framework and an Algorithm”. Ed. by Jude W. Shavlik. Morgan Kaufmann, 1998,
pp. 242–250.
References G. Aguzzi 16 / 20
17. References II
[NNN20] Thanh Thi Nguyen, Ngoc Duy Nguyen, and Saeid Nahavandi. “Deep Reinforcement
Learning for Multiagent Systems: A Review of Challenges, Solutions, and Applications”.
IEEE Trans. Cybern. 50.9 (2020), pp. 3826–3839. DOI: 10.1109/TCYB.2020.2977374.
[MLF07] Laëtitia Matignon, Guillaume J. Laurent, and Nadine Le Fort-Piat. “Hysteretic q-learning :
an algorithm for decentralized reinforcement learning in cooperative multi-agent teams”.
IEEE, 2007, pp. 64–69. DOI: 10.1109/IROS.2007.4399095.
[HKT19] Pablo Hernandez-Leal, Bilal Kartal, and Matthew E. Taylor. “A survey and critique of
multiagent deep reinforcement learning”. Auton. Agents Multi Agent Syst. 33.6 (2019),
pp. 750–797. DOI: 10.1007/s10458-019-09421-1.
[JMK17] Junchen Jin, Xiaoliang Ma, and Iisakki Kosonen. “A stochastic optimization framework
for road traffic controls based on evolutionary algorithms and traffic simulation”. Adv.
Eng. Softw. 114 (2017), pp. 348–360. DOI: 10.1016/j.advengsoft.2017.08.005.
[PL05] Liviu Panait and Sean Luke. “Cooperative Multi-Agent Learning: The State of the Art”.
Auton. Agents Multi Agent Syst. 11.3 (2005), pp. 387–434. DOI:
10.1007/s10458-005-2631-2.
References G. Aguzzi 17 / 20
18. References III
[PMV13] Danilo Pianini, Sara Montagna, and Mirko Viroli. “Chemical-oriented Simulation of
Computational Systems with Alchemist”. Journal of Simulation (2013). ISSN: 1747-7778.
DOI: 10.1057/jos.2012.27.
[HLM15] Jianye Hao, Ho-fung Leung, and Zhong Ming. “Multiagent Reinforcement Social Learning
toward Coordination in Cooperative Multiagent Systems”. ACM Trans. Auton. Adapt. Syst.
9.4 (2015), 20:1–20:20. DOI: 10.1145/2644819.
[Aud+17] Giorgio Audrito et al. “Compositional Blocks for Optimal Self-Healing Gradients”. IEEE
Computer Society, 2017, pp. 91–100. DOI: 10.1109/SASO.2017.18.
References G. Aguzzi 18 / 20
19. Aggregate Computing + Supervised Learning
Ground truth generation is done with simulations
A network is trained with a global view and then
Each agent have the same network
S
10
15
20
15
0
10
25
25
30
S
10
15
20
15
0
10
25
25
30
S
0
10
25
25
30
Appendix G. Aguzzi 19 / 20
20. Aggregate Computing + Evolutionary computing
A population of Network is shared within the entire system
During the simulation, a fitness function is evaluated
Already used in Swarm robotics and automatic design
S
?
?
?
?
?
S
0
10
25
25
30
S
0
10
25
25
30
Appendix G. Aguzzi 20 / 20