Overview

As per the course website, "In this seminar this week we will be looking at hybrid agent architectures that combine reactive and deliberative behaviours. It will involve looking at the underpinning ideas that form this type of architecture and evaluating its relative merits. This will be underlined with a case study that details the deployment of a hybrid architecture and the reasons for its selection."

My Reflection

Overall Reflection

This week was more focused on hybrid agent architectures, as a way of combining both reactive and deliberative architectures and therefore behaviours. The unit only included a reading and a seminar. The latter did not explain much about the topic, but mentioned a good example for hybrid agents which I needed when I was doing the readings. That example was the autonomous delivery robots of Amazon; they have reactive layers and behaviors to avoid running into obstacles, and at the same time they necessarily have deliberative layers and behaviors to navigate the environment and reach the destination.

For the reading, we were required to read 5.2 sub-chapter on Hybrid Agents, from the module's e-book An Introduction to MultiAgent Systems, 2nd. ed. by Wooldridge (2009), which I did, but before that I also read Chapter 4 on Practical Reasoning Agents, and Sub-Chapter 5.1 on Reactive Agents, to have better context before getting into hybrid agents.

Readings Reflection

The readings were overall easy to follow and interesting for me, especially that most of the chapters on practical reasoning and deliberation were relevant to humans as much as machine agents. For example, intentions are defined as "the states of affars that an agent has chosen and committed to," while having an intention and to believing you will bring about it is called intention-belief incosistency, and it is called intention-belief incompleteness whe not believing that the outcome of the intention will be the case. Also, there are degrees of boldness for agents commitment to achieve their intentions, and that different environment types require different intention reconsideration and commitment strategies.

Other than this, a lot of other concepts were mentioned, like that the first planning agent systems were STRIPS system, and how symbloic agents work, given the blocks world example. Then, the text goes on to mention Rodney Brooks, who was interesting for me last week, and his idea of autonomous agents that require to formal knowledge representation, and how this led to the development of reactive agents concept, or what Brooks called subsumption architectures.

Then, the text built on this to reach the idea of hybrid agents and their horizontal and vertical architectures. Horizontal architectures are more straightforward but may require a central controlling module to decide the overall 'attitude' of the agent, which might turn into a bottleneck in the case of all agents but the very simple agents, while vertical architectures are more complex but suitable for a wider array of cases.

What I still needed most was relevant examples from the today's world. For example, I was wondering about the agents that are connected to LLM chatbots and what type do they fall under. What about MCPs? Are they considered as agents, and if yes, what type of agents? Maybe the fact that the e-book is from 2009 has had this effect of absence of relevant up-to-date examples. Hope I will be able to connect those concepts with the latest developments of AI agents down the line of this module.

Wooldridge, M.J. (2009) An Introduction to Multiagent Systems. 2nd edn. Chichester: John Wiley & Sons.