SlideShare a Scribd company logo
Hierarchical
Reinforcement Learning




    David Jardim & Luís Nunes
      ISCTE-IUL 2009/2010
Hierarchical
Reinforcement Learning




    David Jardim & Luís Nunes
      ISCTE-IUL 2009/2010
Outline 1/2
Planning Process
The Problem and Motivation
Reinforcement Learning
Markov Decision Process
Q-Learning
Hierarchical Reinforcement Learning
  Why HRL?
  Approaches
                   3
Outline 2/2
  Semi-Markov Decision Process
  Options
Until Now
Next Step - Simbad
Limitations of HRL
Future Work on HRL
Questions
References

                     4
Planning Process




       5
The Problem and




                                              LEGO_Mindstorms_NXT_mini.jpg
                                              @ http:/
       Motivation




                                                      /lambcutlet.org/images/
LEGO MindStorms Robot with sensors,
actuators and noise

Purpose of collecting “bricks” and assembly
them accordingly to a plan

Decompose the global problem in sub-
problems

Try to solve the problem by implementing
well-known RL and HRL techniques

                    6
Reinforcement Learning

 Computational
 approach to learning       @ R. S. Sutton, Reinforcement Learning: An Introduction
                                               (MIT Press, 1998).


 An agent tries to maximize the reward he receives
 when an action is taken

 Interacts with a complex, uncertain environment

 Learns how to map situations to actions


                        7
Markov Decision Process

 A finite MDP is defined by

   a finite set of states S

   a finite set of actions A


                              @ http://en.wikipedia.org/wiki/Markov_decision_process




                     8
Q-Learning
          [Watkins, C.J.C.H.’89]


Agent with a state set S and action set A.

Performs an action a in order to change its
state.

A reward is provided by the environment.

The goal of the agent is to maximize its
total reward.


             @ http://en.wikipedia.org/wiki/Q-learning
                                9
Why HRL?
Improve the performance

Impossibility to apply RL to problems with
large state/action (curse of dimensionality)

Sub-goals and abstract actions can be used
on different tasks (state abstraction)

Multiple levels of temporal abstraction

Obtain state abstraction

                    10
Approaches
HAMs - Hierarchies of Abstract Machines (Parr
& Russell, 98)

Options - Between MDPs and Semi-MDPs:
Learning, Planning, and Representing Knowledge
at Multiple Temporal Scales (Sutton, Precup &
Singh, 99)

MAXQ Value Function Decomposition (Dietterich,
2000)

Discovering Hierarchy in RL with HEXQ (Hengst,
2002)
                    11
Approaches
HAMs - Hierarchies of Abstract Machines (Parr
& Russell, 98)

Options - Between MDPs and Semi-MDPs:
Learning, Planning, and Representing Knowledge
at Multiple Temporal Scales (Sutton, Precup &
Singh, 99)

MAXQ Value Function Decomposition (Dietterich,
2000)

Discovering Hierarchy in RL with HEXQ (Hengst,
2002)
                    11
Semi-Markov Decision
      Process
An SMDP consists of

  A set of states S

  A set of actions A

  An expected cumulative discounted reward

  A well-defined joint distribution of the
  next state and transit time


                      12
Options
       [Sutton, Precup & Singh’99]


An Option is defined by

  A policy ∏: SxA ➞ [0,1]

  A termination condition β: S^+ →[0,1]

  And an initiation set I⊆S

Its hierarchical and used to reach sub-goals



                      13
Until Now




O1

     O2




              14
Until Now
                                                                     St

Steps                                      Steps




                  Episodes                              Episodes



        @ Sutton, Precup & Singh’99                @ My Simulation



                                      15
Next Step - Simbad
Java 3D Robot Simulator

3D visualization and sensing

Range Sensor: sonars and IR

Contact Sensor: bumpers               @ http://simbad.sourceforge.net/



Will allow us to simulate and learn first, and then
transfer the learning to our LEGO MindStorm


                        16
Limitations of HRL

Effectiveness of these ideas on large and
complex continuous control tasks

Sub-goals are assigned manually

Some of the existing algorithms only work
well for the problem which they were
designed to solve


                   17
Future Work on HRL

Automated discovery of state abstraction

Find the best automated way to discovery
sub-goals to associate with Options

Obtain a long lived learning agent that faces
a continued series of tasks and keep evolving



                   18
Questions?




    19
Questions?




    19
References

R. S. Sutton, Reinforcement Learning: An Introduction (MIT Press, 1998).

R. Parr and S. Russell. Reinforcement learning with hierarchies of machines. In Advances in
Neural Information Processing Systems: Proceedings of the 1997 Conference, Cambridge,
MA, 1998. MIT Press.

R. S. Sutton, D. Precup, and S. Singh. Between mdps and semi-mdps: A framework for
temporal abstraction in reinforcement learning. Artificial Intelligence, 112:181–211, 1999.

T. G. Dietterich. Hierarchical reinforcement learning with the MAXQ value function
decomposition. Journal of Artificial Intelligence Research, 13:227–303, 2000.

B. Hengst. Discovering hierarchy in reinforcement learning with hexq. In Maching Learning:
Proceedings of the Nineteenth International Conference on Machine Learning, 2002.




                                         20

More Related Content

What's hot (20)

PDF
An introduction to deep reinforcement learning
Big Data Colombia
 
PPT
Reinforcement Learning Q-Learning
Melaku Eneayehu
 
PDF
Markov decision process
Hamed Abdi
 
PPTX
Reinforcement Learning
DongHyun Kwak
 
PDF
Reinforcement learning
DongHyun Kwak
 
PDF
Reinforcement Learning 4. Dynamic Programming
Seung Jae Lee
 
PDF
A brief overview of Reinforcement Learning applied to games
Thomas da Silva Paula
 
PDF
Deep Reinforcement Learning: Q-Learning
Kai-Wen Zhao
 
PDF
Reinforcement Learning
Muhammad Iqbal Tawakal
 
PDF
Deep learning - A Visual Introduction
Lukas Masuch
 
PDF
Deep Reinforcement Learning
MeetupDataScienceRoma
 
PDF
Deep Q-Learning
Nikolay Pavlov
 
PPTX
Reinforcement learning
Ding Li
 
PPTX
Unsupervised learning (clustering)
Pravinkumar Landge
 
PDF
Multi-armed Bandits
Dongmin Lee
 
PPTX
Machine Learning and Real-World Applications
MachinePulse
 
PDF
Reinforcement Learning
CloudxLab
 
PDF
SVM Algorithm Explained | Support Vector Machine Tutorial Using R | Edureka
Edureka!
 
PDF
Deep learning - what is it and why now?
Natalia Konstantinova
 
PPTX
Intro to Deep Reinforcement Learning
Khaled Saleh
 
An introduction to deep reinforcement learning
Big Data Colombia
 
Reinforcement Learning Q-Learning
Melaku Eneayehu
 
Markov decision process
Hamed Abdi
 
Reinforcement Learning
DongHyun Kwak
 
Reinforcement learning
DongHyun Kwak
 
Reinforcement Learning 4. Dynamic Programming
Seung Jae Lee
 
A brief overview of Reinforcement Learning applied to games
Thomas da Silva Paula
 
Deep Reinforcement Learning: Q-Learning
Kai-Wen Zhao
 
Reinforcement Learning
Muhammad Iqbal Tawakal
 
Deep learning - A Visual Introduction
Lukas Masuch
 
Deep Reinforcement Learning
MeetupDataScienceRoma
 
Deep Q-Learning
Nikolay Pavlov
 
Reinforcement learning
Ding Li
 
Unsupervised learning (clustering)
Pravinkumar Landge
 
Multi-armed Bandits
Dongmin Lee
 
Machine Learning and Real-World Applications
MachinePulse
 
Reinforcement Learning
CloudxLab
 
SVM Algorithm Explained | Support Vector Machine Tutorial Using R | Edureka
Edureka!
 
Deep learning - what is it and why now?
Natalia Konstantinova
 
Intro to Deep Reinforcement Learning
Khaled Saleh
 

Viewers also liked (20)

PDF
HRL: Learning Subgoals and State Abstraction
David Jardim
 
PDF
Generalized Reinforcement Learning
Po-Hsiang (Barnett) Chiu
 
PDF
Htn in videogames
Andrea Tucci
 
PPTX
My culture assignment
Shawn Cap
 
PDF
World Cities Culture Report - 2015
WorldCitiesCultureForum
 
PPT
Culture and community report
Evelyn Avila
 
PPTX
Popular Culture Assignment 1 5
Todd Messner
 
DOCX
Cross clulture uk
badalsolapurwala01
 
PPTX
Culture and communication
BelaCELT
 
PPTX
Hard to measure: Company Culture Report Summary Findings
James Bridgman
 
PPT
Intercultural communication.sal july 2010
sun1you
 
PPTX
Bangladesh culture
arifplus
 
PPTX
Axact - OB report on Organizational Culture
Fayaz T
 
PPTX
Culture and communication
prezantimedetyra
 
PPTX
My Cultural Project Intro
Chris Allen
 
PPTX
Culture and Nonverbal Communication in Italy
SNash53328
 
PPTX
Relationships between communication and culture
xochitlfaro
 
PPTX
High context cross culture communication of india
Vikram M. Nimbal
 
PPTX
Communication culture and context
Surbhi Parashar
 
HRL: Learning Subgoals and State Abstraction
David Jardim
 
Generalized Reinforcement Learning
Po-Hsiang (Barnett) Chiu
 
Htn in videogames
Andrea Tucci
 
My culture assignment
Shawn Cap
 
World Cities Culture Report - 2015
WorldCitiesCultureForum
 
Culture and community report
Evelyn Avila
 
Popular Culture Assignment 1 5
Todd Messner
 
Cross clulture uk
badalsolapurwala01
 
Culture and communication
BelaCELT
 
Hard to measure: Company Culture Report Summary Findings
James Bridgman
 
Intercultural communication.sal july 2010
sun1you
 
Bangladesh culture
arifplus
 
Axact - OB report on Organizational Culture
Fayaz T
 
Culture and communication
prezantimedetyra
 
My Cultural Project Intro
Chris Allen
 
Culture and Nonverbal Communication in Italy
SNash53328
 
Relationships between communication and culture
xochitlfaro
 
High context cross culture communication of india
Vikram M. Nimbal
 
Communication culture and context
Surbhi Parashar
 
Ad

Similar to Hierarchical Reinforcement Learning (20)

PDF
Matineh Shaker, Artificial Intelligence Scientist, Bonsai at MLconf SF 2017
MLconf
 
PDF
Deep RL for Autonomous Driving exploring applications Cognitive vehicles 2019
Ravi Kiran B.
 
PDF
cs330_2021_lifelong_learning.pdf
Kuan-Tsae Huang
 
PDF
Deep Learning for Real-Time Atari Game Play Using Offline Monte-CarloTree Sear...
Willy Marroquin (WillyDevNET)
 
PPTX
Get hands-on with Explainable AI at Machine Learning Interpretability(MLI) Gym!
Sri Ambati
 
PDF
Entity Summarization with User Feedback (ESWC 2020)
Qingxia Liu
 
PDF
An exhaustive survey of reinforcement learning with hierarchical structure
eSAT Journals
 
PDF
State representation learning for control: an overview
Natalia Díaz Rodríguez
 
PDF
Toward unified framework and symbolic decision making - Berkeley LLM AI Agent...
VincentLui15
 
PPTX
Unlocking Exploration: Self-Motivated Agents Thrive on Memory-Driven Curiosity
Hung Le
 
PPTX
Human Level Artificial Intelligence
Rahul Chaurasia
 
PDF
10 1 planning, acting, learning
Tianlu Wang
 
PDF
Deep reinforcement learning from scratch
Jie-Han Chen
 
PDF
Presentation v2
MehrnooshV
 
DOC
Hibridization of Reinforcement Learning Agents
butest
 
PDF
Reinforcement Learning (DLAI D7L2 2017 UPC Deep Learning for Artificial Intel...
Universitat Politècnica de Catalunya
 
PPT
Scheduling And Htn
ahmad bassiouny
 
PDF
Location Prediction Under Data Sparsity
James McInerney
 
PDF
MILA DL & RL summer school highlights
Natalia Díaz Rodríguez
 
PDF
Introduction of Deep Reinforcement Learning
NAVER Engineering
 
Matineh Shaker, Artificial Intelligence Scientist, Bonsai at MLconf SF 2017
MLconf
 
Deep RL for Autonomous Driving exploring applications Cognitive vehicles 2019
Ravi Kiran B.
 
cs330_2021_lifelong_learning.pdf
Kuan-Tsae Huang
 
Deep Learning for Real-Time Atari Game Play Using Offline Monte-CarloTree Sear...
Willy Marroquin (WillyDevNET)
 
Get hands-on with Explainable AI at Machine Learning Interpretability(MLI) Gym!
Sri Ambati
 
Entity Summarization with User Feedback (ESWC 2020)
Qingxia Liu
 
An exhaustive survey of reinforcement learning with hierarchical structure
eSAT Journals
 
State representation learning for control: an overview
Natalia Díaz Rodríguez
 
Toward unified framework and symbolic decision making - Berkeley LLM AI Agent...
VincentLui15
 
Unlocking Exploration: Self-Motivated Agents Thrive on Memory-Driven Curiosity
Hung Le
 
Human Level Artificial Intelligence
Rahul Chaurasia
 
10 1 planning, acting, learning
Tianlu Wang
 
Deep reinforcement learning from scratch
Jie-Han Chen
 
Presentation v2
MehrnooshV
 
Hibridization of Reinforcement Learning Agents
butest
 
Reinforcement Learning (DLAI D7L2 2017 UPC Deep Learning for Artificial Intel...
Universitat Politècnica de Catalunya
 
Scheduling And Htn
ahmad bassiouny
 
Location Prediction Under Data Sparsity
James McInerney
 
MILA DL & RL summer school highlights
Natalia Díaz Rodríguez
 
Introduction of Deep Reinforcement Learning
NAVER Engineering
 
Ad

Recently uploaded (20)

PDF
Blockchain Transactions Explained For Everyone
CIFDAQ
 
PDF
July Patch Tuesday
Ivanti
 
PDF
Using FME to Develop Self-Service CAD Applications for a Major UK Police Force
Safe Software
 
PDF
HCIP-Data Center Facility Deployment V2.0 Training Material (Without Remarks ...
mcastillo49
 
PDF
Presentation - Vibe Coding The Future of Tech
yanuarsinggih1
 
PPTX
Webinar: Introduction to LF Energy EVerest
DanBrown980551
 
PPTX
OpenID AuthZEN - Analyst Briefing July 2025
David Brossard
 
PDF
"AI Transformation: Directions and Challenges", Pavlo Shaternik
Fwdays
 
PPTX
Q2 FY26 Tableau User Group Leader Quarterly Call
lward7
 
PPTX
COMPARISON OF RASTER ANALYSIS TOOLS OF QGIS AND ARCGIS
Sharanya Sarkar
 
PDF
Building Real-Time Digital Twins with IBM Maximo & ArcGIS Indoors
Safe Software
 
PDF
IoT-Powered Industrial Transformation – Smart Manufacturing to Connected Heal...
Rejig Digital
 
PDF
Newgen Beyond Frankenstein_Build vs Buy_Digital_version.pdf
darshakparmar
 
PDF
Exolore The Essential AI Tools in 2025.pdf
Srinivasan M
 
PDF
NewMind AI - Journal 100 Insights After The 100th Issue
NewMind AI
 
PDF
Transcript: New from BookNet Canada for 2025: BNC BiblioShare - Tech Forum 2025
BookNet Canada
 
PDF
The Rise of AI and IoT in Mobile App Tech.pdf
IMG Global Infotech
 
PDF
Mastering Financial Management in Direct Selling
Epixel MLM Software
 
PDF
Agentic AI lifecycle for Enterprise Hyper-Automation
Debmalya Biswas
 
PDF
Chris Elwell Woburn, MA - Passionate About IT Innovation
Chris Elwell Woburn, MA
 
Blockchain Transactions Explained For Everyone
CIFDAQ
 
July Patch Tuesday
Ivanti
 
Using FME to Develop Self-Service CAD Applications for a Major UK Police Force
Safe Software
 
HCIP-Data Center Facility Deployment V2.0 Training Material (Without Remarks ...
mcastillo49
 
Presentation - Vibe Coding The Future of Tech
yanuarsinggih1
 
Webinar: Introduction to LF Energy EVerest
DanBrown980551
 
OpenID AuthZEN - Analyst Briefing July 2025
David Brossard
 
"AI Transformation: Directions and Challenges", Pavlo Shaternik
Fwdays
 
Q2 FY26 Tableau User Group Leader Quarterly Call
lward7
 
COMPARISON OF RASTER ANALYSIS TOOLS OF QGIS AND ARCGIS
Sharanya Sarkar
 
Building Real-Time Digital Twins with IBM Maximo & ArcGIS Indoors
Safe Software
 
IoT-Powered Industrial Transformation – Smart Manufacturing to Connected Heal...
Rejig Digital
 
Newgen Beyond Frankenstein_Build vs Buy_Digital_version.pdf
darshakparmar
 
Exolore The Essential AI Tools in 2025.pdf
Srinivasan M
 
NewMind AI - Journal 100 Insights After The 100th Issue
NewMind AI
 
Transcript: New from BookNet Canada for 2025: BNC BiblioShare - Tech Forum 2025
BookNet Canada
 
The Rise of AI and IoT in Mobile App Tech.pdf
IMG Global Infotech
 
Mastering Financial Management in Direct Selling
Epixel MLM Software
 
Agentic AI lifecycle for Enterprise Hyper-Automation
Debmalya Biswas
 
Chris Elwell Woburn, MA - Passionate About IT Innovation
Chris Elwell Woburn, MA
 

Hierarchical Reinforcement Learning

  • 1. Hierarchical Reinforcement Learning David Jardim & Luís Nunes ISCTE-IUL 2009/2010
  • 2. Hierarchical Reinforcement Learning David Jardim & Luís Nunes ISCTE-IUL 2009/2010
  • 3. Outline 1/2 Planning Process The Problem and Motivation Reinforcement Learning Markov Decision Process Q-Learning Hierarchical Reinforcement Learning Why HRL? Approaches 3
  • 4. Outline 2/2 Semi-Markov Decision Process Options Until Now Next Step - Simbad Limitations of HRL Future Work on HRL Questions References 4
  • 6. The Problem and LEGO_Mindstorms_NXT_mini.jpg @ http:/ Motivation /lambcutlet.org/images/ LEGO MindStorms Robot with sensors, actuators and noise Purpose of collecting “bricks” and assembly them accordingly to a plan Decompose the global problem in sub- problems Try to solve the problem by implementing well-known RL and HRL techniques 6
  • 7. Reinforcement Learning Computational approach to learning @ R. S. Sutton, Reinforcement Learning: An Introduction (MIT Press, 1998). An agent tries to maximize the reward he receives when an action is taken Interacts with a complex, uncertain environment Learns how to map situations to actions 7
  • 8. Markov Decision Process A finite MDP is defined by a finite set of states S a finite set of actions A @ http://en.wikipedia.org/wiki/Markov_decision_process 8
  • 9. Q-Learning [Watkins, C.J.C.H.’89] Agent with a state set S and action set A. Performs an action a in order to change its state. A reward is provided by the environment. The goal of the agent is to maximize its total reward. @ http://en.wikipedia.org/wiki/Q-learning 9
  • 10. Why HRL? Improve the performance Impossibility to apply RL to problems with large state/action (curse of dimensionality) Sub-goals and abstract actions can be used on different tasks (state abstraction) Multiple levels of temporal abstraction Obtain state abstraction 10
  • 11. Approaches HAMs - Hierarchies of Abstract Machines (Parr & Russell, 98) Options - Between MDPs and Semi-MDPs: Learning, Planning, and Representing Knowledge at Multiple Temporal Scales (Sutton, Precup & Singh, 99) MAXQ Value Function Decomposition (Dietterich, 2000) Discovering Hierarchy in RL with HEXQ (Hengst, 2002) 11
  • 12. Approaches HAMs - Hierarchies of Abstract Machines (Parr & Russell, 98) Options - Between MDPs and Semi-MDPs: Learning, Planning, and Representing Knowledge at Multiple Temporal Scales (Sutton, Precup & Singh, 99) MAXQ Value Function Decomposition (Dietterich, 2000) Discovering Hierarchy in RL with HEXQ (Hengst, 2002) 11
  • 13. Semi-Markov Decision Process An SMDP consists of A set of states S A set of actions A An expected cumulative discounted reward A well-defined joint distribution of the next state and transit time 12
  • 14. Options [Sutton, Precup & Singh’99] An Option is defined by A policy ∏: SxA ➞ [0,1] A termination condition β: S^+ →[0,1] And an initiation set I⊆S Its hierarchical and used to reach sub-goals 13
  • 15. Until Now O1 O2 14
  • 16. Until Now St Steps Steps Episodes Episodes @ Sutton, Precup & Singh’99 @ My Simulation 15
  • 17. Next Step - Simbad Java 3D Robot Simulator 3D visualization and sensing Range Sensor: sonars and IR Contact Sensor: bumpers @ http://simbad.sourceforge.net/ Will allow us to simulate and learn first, and then transfer the learning to our LEGO MindStorm 16
  • 18. Limitations of HRL Effectiveness of these ideas on large and complex continuous control tasks Sub-goals are assigned manually Some of the existing algorithms only work well for the problem which they were designed to solve 17
  • 19. Future Work on HRL Automated discovery of state abstraction Find the best automated way to discovery sub-goals to associate with Options Obtain a long lived learning agent that faces a continued series of tasks and keep evolving 18
  • 22. References R. S. Sutton, Reinforcement Learning: An Introduction (MIT Press, 1998). R. Parr and S. Russell. Reinforcement learning with hierarchies of machines. In Advances in Neural Information Processing Systems: Proceedings of the 1997 Conference, Cambridge, MA, 1998. MIT Press. R. S. Sutton, D. Precup, and S. Singh. Between mdps and semi-mdps: A framework for temporal abstraction in reinforcement learning. Artificial Intelligence, 112:181–211, 1999. T. G. Dietterich. Hierarchical reinforcement learning with the MAXQ value function decomposition. Journal of Artificial Intelligence Research, 13:227–303, 2000. B. Hengst. Discovering hierarchy in reinforcement learning with hexq. In Maching Learning: Proceedings of the Nineteenth International Conference on Machine Learning, 2002. 20

Editor's Notes

  • #2: Boa tarde. Estou cá para vós falar sobre o meu projecto de dissertação, que insere-se na área da aprendizagem hierárquica por reforço.
  • #4: Durante esta apresentação ou abordar o problema e a motivação para o mesmo, definir a aprendizagem por reforço e a sua estrutura matemática. Falar um pouco sobre o QLearning e posteriormente aprofundar o tema da Aprendizagem por Reforço Hierárquica.
  • #5: Mostrar algum do trabalho desenvolvido até ao momento, e definir quais são os próximos passos.
  • #7: Pretende-se simular um robô que tem o objectivo de buscar tijolos e dispôr os mesmos de acordo com um plano. Vou tentar dividir o problema em várias tarefas, por ex: como encontrar o tijolo, empurrar o tijolo... Numa fase seguinte, efectuar o mesmo com um robô real num cenário real. A questão é até que ponto, as técnicas conhecidas de aprendizagem por reforço e aprendizagem por reforço hierarquico nos podem levar à resolução do problema começando por uma simplificação e adicionar complexidade progressivamente.
  • #8: A aprendizagem por reforço na área computacional, consiste, numa abordagem à aprendizagem, onde um agente ao executar uma acção recebe uma recompensa e altera o seu estado. Ao longo do tempo essa recompensa vai permitir mapear estados para acções e criar uma política. Essa política vai permitir ao agente resolver o problema a que foi proposto.
  • #9: Se uma tarefa em aprendizagem por reforço possui um conjunto de estados e de acções finitos, então podemos afirmar que essa mesma tarefa é um Markov Decision Process. Para qualquer estado e acção, a probabilidade do estado seguinte ocorrer é definido pela 1ª equação. Consiste na recompensa imediata após transitar para o estado seguinte com a probabilidade definida anteriormente
  • #10: Foi uma descoberta muito importante para a área, pode ser vista como um agente que escolhe uma acção a partir de uma política, de seguida executa essa mesma acção, recebe uma recompensa e transita para um novo estado actualizando a qualidade da acção executada no estado correspondente.
  • #11: A aprendizagem por reforço possui algumas limitações, consoante a complexidade do problema pode-se revelar impracticável a sua aprendizagem. Quanto mais complexo maior é o conjunto de estados e de acções. A aprendizagem por reforço pode ser utilizada para acelerar o processo de aprendizagem, diminuir a quantidade de recursos necessários (memória), reutilizar a aprendizagem adquirida em diferentes tarefas (state abstraction). Dessa forma tornar possível a resolução de alguns problemas impossíveis.
  • #12: HAMs - Através de uma hierarquia de máquinas de estados finitos, organizada por ordem crescente de complexidade, onde as máquinas de topo são compostas pelas máquinas subjacentes até chegarmos às máquinas que executam as acções primitivas. MAXQ - Trata o problema como um conjunto de problemas QLearning simultâneos. Consegue decompôr a política de forma a aproveitar partes que se repetem. Implementa vários tipos de state abstraction. HEXQ - Tenta decompôr um MDP ao dividir o espaço dos estados em regiões sub-MDP aninhadas e posteriormente tenta resolver o problema para cada uma das regiões.
  • #13: Os Semi-Markov Decision Process são considerados um tipo especial de MDP, apropriados para modelar sistemas de eventos discretos em tempo continuo. A grande diferença é que neste caso as acções podem levar quantidades de tempo variáveis de forma a modelarem acções temporalmente alargadas.
  • #14: Esta foi a abordagem escolhida como base para o trabalho a desenvolver, porque em relação às outras abordagens, esta revelou-se a mais flexível em termos de problemas em que pode ser aplicada. Politica - Consiste no conjunto de acções primitivas que definem a acção composta Condição de terminação - A probabilidade é igual a 1 quando o agente muda de sala por ex. Consiste no conjunto de estados onde a Option pode ser iniciada (uma sala). Criação da OPTION atráves de QLearning
  • #15: De forma a obter uma base sólida de conhecimento acerca das OPTIONS, tentei replicar o estudo efectuado pelos autores, e esta foi a minha implementação. Na 1ª imagem temos uma representação de uma OPTION, que não é nada mais que uma acção composta. E na 2ª imagem podemos ver os estados onde o agente decidiu que era mais vantajoso executar uma OPTION ou uma acção primitiva.
  • #16: Os resultados obtidos foram muito próximos da simulação efectuada pelos autores, como podem ver pela comparação entre os dois gráficos. Neste momento está criada a base que vai ser utilizada para efectuar testes e tentar resolver o problema, quem sabe até melhorar estes resultados.
  • #17: O passo seguinte é utilizar uma plataforma de simulação de robôs com diversas funcionalidades, e tentar resolver o problema proposto do robô construtor utilizando os conhecimentos adquiridos na implementação das OPTIONS. Quando esse objectivo for alcançado, então começarei a trabalhar para transferir a aprendizagem adquirida na simulação para o robô real e fazer com que ele resolva o problema aprendido.
  • #18: É claro que a aprendizagem por reforço hierárquica não é perfeita, porque por enquanto apenas consegue ser aplicada a problemas de pequenas ou médias dimensões, e os sub-objectivos são atribuidos de forma manual pelo programador. Também a maior parte dos algorítmos desenvolvidos apenas funcionam como devem de ser para os problemas que foram pensados incialmente.
  • #19: Existe muito trabalho futuro nesta área e as possibilidades são imensas. Pela investigação efectuada o maior desafio consiste na descoberta automática de estruturas hierárquicas, como a descoberta dos sub-objectivos de forma a que o agente divida o problema em vários problemas mais simples, e fazer com que o agente atráves do processo de aprendizagem seja capaz de responder a novos desafios e evoluir de acordo.
  • #20: Obrigado, agora se alguem tiver questões que eu possa responder.