Font Size: 

In today's modern world, artificial intelligence and more specifically the autonomous agents are gradually becoming more and more a popular topic. Life is quite easier when someone or something can do the work for us. However, to be able to perform any useful task or even perform it efficiently, the autonomous agents need to be able to make the correct decisions in any given time. The ability to make such decisions that the agent (not only artificial but any kind of being that we can call an agent) will fulfil its goals, is called rationality. More accurately, rationality is optimal decision making, given one's believes (internal knowledge) and the available observations about the world. Even today, there are numerous systems already implemented in real-life applications that are capable of rational decision making in tasks with various degrees of complexity. These systems are mostly based on an inference mechanism and a knowledge database. Therefore, they use exact, logical principles to make decisions. But what about systems that don't use logic as such and are less exact - are they capable of performing rational behaviour? And what would be the advantages of such systems? Neural networks seem to be a very promising example of implicit decision-making systems, so perhaps we can find other biologically inspired principles that can be used for this purpose.

In my work, I tested the idea of using emotionally inspired models for making decisions in control systems of autonomous agents. My first task was to research what has already been done in this area and most importantly, to find out whether the emotions can be considered rational in any way. I selected several emotional models that had already been proposed such as the Fungus Eater ([1], [2]). Based on the knowledge gained by my literature research I attempted to devise my own emotional model so that I could test the idea of using emotions for optimal decision-making.

My model uses three simple emotion equivalents (emotions in my model need to be understood only as analogies to real emotions - therefore the term "emotion" in this abstract should not be understood as real emotion but merely as something inspired by real emotions). These emotions are generated when the agent encounters an object and based on them the agent can create an attitude towards each object in its surrounding. The attitude is determined by several parameters that specify for example how much wants the agent to examine or interact with the object or how much it wants to avoid the object. To generate these responses, the emotional system uses an artificial neural network. That means it can learn the correct responses either from training examples, if we want to prepare the system for the environment where it will be applied, or on the fly after the agent has been put into the environment. The latter uses the feedback transformed to (correct) emotional responses that the agent receives when it interacts with an object (e.g. refilling its batteries from an electric socket gives a positive response, which is equal to "joy").

After I had devised the emotional model, I implemented it into a control system for an agent. Based on the attitudes created by the emotional model and the agent's internal state, the control system makes decisions about what should be the agent's next short-term goal (e.g. when batteries are low and there is a socket nearby, go for the socket and plug into it). The agent's internal state consists of "needs" (e.g. the state of batteries is represented by the need of electricity, the lower is the energy in batteries the higher is the need). Besides the basic needs required for existence of the agent, it is also possible to add needs to control the agent's behaviour. For example if the agent should perform a specific task, it can have a need that will increase over time if the agent is not fulfilling the task. The need will be then decreased if the agent performs the task. The agent prefers interacting with objects related to the higher needs, so it will try to satisfy the need with the higher value (if it can be satisfied at the moment).

My last task was to implement and test both the emotional model and the control system in a simple multi-agent based computer simulation. The main goal of the agents in this simulation was to survive in a mostly unknown environment. The agents needed to learn how they should react on the objects in the environment (e.g. what they can eat and what objects they should avoid).

Considering the results of the simulation, I can say that the agents behaved rationally. This means that the control system with an emotional model really works and it can be used for rational autonomous agents. Based on my work, I can conclude that the emotional models show a great potential in the area of artificial intelligence.

References:
[1] Toda, M. The Urge Theory of Emotion and Social Interaction: Chapter 6 and 7, 1998.
[2] Ruebenstrunk, G. Emotional Computers, 1998. http://www.ruebenstrunk.de/emeocomp/content.HTM (27.4. 2010)
[3] Velasquez, J.D. When robots weep: Emotional memories and decision-making. MIT, www.aaai.org, 1998.