Personal tools
You are here: Home Member Resources info NIAD&R Info Agents’ Adaptation, Learning and Emotions

Agents’ Adaptation, Learning and Emotions

Agents’ intelligent processes mostly rely on learning capabilities and sophisticated architectures. Through this research line we aim at studying both agent and multi-agent learning on one hand, and emotion-like features driven architectures, on the other hand.

Researchers:

 Eugénio Oliveira, Luís Paulo Reis, Luís Nunes, Luís Sarmento, Vasco Vinhas, Hugo Gravato Marques

Coordinator:

Eugénio Oliveira

Research direction:

Agents’ intelligent processes mostly rely on learning capabilities and sophisticated architectures. Through this research line we aim at studying both agent and multi-agent learning on one hand, and emotion-like features driven architectures, on the other hand.

(i) Agents and Multi-agent learning. The main goal of this research issue is to find an answer to the following question: ”(How) can several di?erent, heterogeneous, Learning Agents improve their performance by exchanging information during their own learning process?”. We are also researching on computational learning methods for bioinformatics.

(ii) Emotion-based agents’ architecture. Through this research issue we would like to answer another important question: ”Will it be possible to escape from usual utility-based decision functions, by using emotion-like features, in what decision-making for autonomous agents as well as for teams of Agents is concerned?”

Agents and Multi-Agent Learning

Research goals:

1- (How) can several different Learning Agents improve their performance by exchanging information during their own learning process?”;

2- Concerning Inductive Logic Programming algorithms applied to Bio-Informatics, how to apply Inductive Logic Programming (ILP) to the Protein Folding problem (collaboration with REQUIMTE); 3- How to apply ILP to analysis of genomic sequences (collaboration with IBMC).

Recent work (2006):
  • Our studies on the effects of communication during learning in teams of agents that use different learning algorithms have been concluded. These studies were based in experiments in different scenarios: the Predator-Prey domain, a Traffic Control simulation based on real data and a Load-Balance simulation.
  • During the above mentioned studies we have identified several weaknesses of the process, proposed and tested new solutions. The weaknesses are mainly related to the synchronization of information exchange in a team and to the integration of advice from peers using different learning algorithms. A PhD thesis on this subject was successfully concluded during 2006.
  • We are using ILP in the prediction of secondary protein structure based on the primary structure and properties of the amino-acids. We also use ILP to find rules that predict the location of helices and beta-sheets based on the linear sequence of residues of each protein. Concerning the problem of genomic sequence analysis we are addressing, we want to ”explain” the behaviour of disjoint exons based on their sequence of both bases surrounding introns.
  • We have developed a program to access periodically the Dunbrac/Pixies Web page to obtain the most recent list of proteins with very low homology. With that list the program fetches the protein sequences at the PDB repository. item We have also developed the basic background knowledge useful to tacle the Protein Folding problem with ILP.
  • We developed pre-processing analysis to the sequences data that allow us to substantially reduce the number of hypotheses during the ILP execution stage. These preprocessing analysis is based on the computation of individual and pair propensities.
  • We developed the background knowledge for the genomic sequence analysis.
Future work: 
  • We intend to apply ILP to 3 sub-problems associated with the prediction of the secondary structure of proteins: i) identify when an alpha-helix starts; ii) predict when an alpha-helix ends and; iii) predict the size of an alpha-helix.
  • We will also apply ILP to the genomic sequence analysis problem.

Emotion-like based Agents

Research goals:

1- Modelling and specifying an emotion based agent architecture;

2- tactical models for coordinating a small team of emotion-based agents

Recent work:
  • While our past work in this area mainly concerned the refinement of basic concepts of the emotion-based agent architecture, the work in 2006 was focused on providing a more appropriate formalization of the new agent architecture. We have thus proposed an extension to the BDI architecture capable of supporting Artificial Emotions, named Emotional-BDI Architecture. This architecture is original in the sense that is designed to extend the BDI architecture with explicit representation of two important concepts that are related to emotional processing: Capabilities and Resources. This is achieved while trying to keeping the same logic formalisms already in place for the BDI architecture.
  • Through the ”Emotional-BDI model”, in which we added new components for managing resources and capabilities, as well as for managing the activation of emotions; second, we extended Rao and George?’s logic, by adding new modal operators which describe the new components present in the Emotional-BDI model. Using this new "E_BDI" logic, we model the activation conditions of three emotions and the effects that their presence has in the agent behaviour.
  • We have introduced the ”E_BDI logic”, by defining its syntax, semantics and describing the properties of the modal operators. We also presented some results about the interaction between time and actions. The formal specification of the activation and effects of the following set of emotions: fear, anxiety and self-confidence have been developed.
  • In 2006 we were able to implement a tactical model to coordinate small teams of emotion-based agents (here, agents are firefighters that try to control a forest fire). In the present model, there are two coordination levels: global coordination and local coordination. Global coordination uses predefined tactics (based on real firefighting tactics) that specify the team overall approach to fire at a high-level. New tactics can be added and it is also possible to extend previously defined tactics. A leader agent evaluates the scenario (using its own perception and information communicated by other agents) and based on the current tactic it assigns high-level tasks to the other agents. Local coordination is necessary to carry out these high-level tasks effectively. Agents that share the same tasks use perception and predefined rules in order to cooperate without using communication. We tested an implementation of this model by experimenting several firefighting tactics in two scenarios. We have demonstrated that, like in reality, different fire scenarios require different firefighting tactics in order to minimize fire damage. Additionally, the tactics that performed best in the tested scenarios are the ones that we were expecting according to firefighting theory.
  • A Master thesis on the Tactics for Emotion-based Agent teams was successfully submitted.
Current and future work:
  • To finalize and submit a Master thesis on the E_BDI agent architecture model.
  • Future work concerning emotion-based agent teams will include performing more experiments in other types of scenarios. We need to enhancing and validating the simulator in order to give more credibility to our experiments. After this, we are planning to use machine-learning algorithms over the the experimental results for trying to discover rules for tactic selection. At this stage, we are also ready to take advantageous of the the agents emotional mechanism for making higher-level decisions, such as tactical decisions, and to study the effect of emotions in the overall team performance.

[Go back]

Document Actions