Formal Models and Computational Limitations

It is interesting to note that affect defined as the positiveness versus the negativeness of a situation (e.g., Gasper & Clore, 2002) is actually a very useful abstraction in the context of Reinforcement Learning. It can be used in many ways, as has been shown in this thesis. However, we have to be careful, again, about conclusions drawn from computational experiments, specifically related to the meaning of the modeled concepts. It can be argued that the way we model affect is quite limited, which is most certainly the case considering the wide variety of emotions and moods that exist in humans. In relation to Reinforcement Learning this definition (and our derived definition of artificial affect) might be adequate, but this does not mean that we have modeled affect in its full glory, or that we can conclude anything about affect in general. Therefore, our psychology- related claims and conclusions have to be interpreted in the context of Reinforcement Learning and instrumental conditioning. Our conclusions are about existence proofs of relations, for they appear beneficial to artificial agents that learn based on different computational models of instrumental conditioning (the versions of RL used in Chapters 3, 4 and 6). As such, they are relevant to experimental psychology. Experimental psychology has difficulties explaining the mechanisms behind relations. In this context, the mechanisms presented in this research are potential candidates that support relations between affect and learning found in the psychological literature. The conclusions should not be carried further than that.

Concrete computer science related results include the control of learning parameters in artificial learning methods by means of abstractions of concepts borrowed from psychology. More specific, artificial affect has successfully been used to control exploration versus exploitation, and affect has been used as reinforcement in an interactive learning setup with a human in the loop. It is very well possible to use affect in a broader sense than the one studied in this thesis. For example, it is interesting to research how affect can be used to control the search through a solution space, as this is also a process of exploration (random jumps, multiple start positions) versus exploitation (hill-climbing). Further,

arousal, the part of affect that defines the activity or action readiness of the

organism—a part we have ignored completely in this thesis—can be modeled and then used to control other parameters. These parameters could be related to the amount of energy available to the agent. Such parameters include the likelihood of acting in the first place and the depth of the thought process.

As mentioned in the previous paragraphs, computational models are limited in their ability to conclude about natural phenomena. This issue has been dealt with related to emotion modeling in Chapter 7. We have shown that it is useful, in fact critical, to use formal models of emotion at an architectural level to advance emotion theory. The analysis has been focused on cognitive appraisal theory, explaining emotions as a result of the subjective evaluation of events in the context of beliefs, desires, and intentions of an agent (being natural or artificial). Our analysis showed that with the formal notation we developed it becomes easier to evaluate whether unexpected behavior resulting from a computational model is due to errors in the computational model or errors in the theory. This is an important issue, as computational models of emotion tend to get very complex and are inspired by many psychological theories (see, e.g., the impressive agent models by Gratch and Marsella, 2004 or Baars and Franklin, 2003). We have further shown that the formal notation can be used to integrate different cognitive appraisal theories, an important issue in the advancement of appraisal theory (Wehrle & Scherer, 2001).

A very valid argument that could be put forward at this point is that we haven’t formally described the affect-learning relations studied in Chapter 3 to 6, and as such can not really draw strong conclusions from these studies. We can say two things about this.

First, we did not formally represent the relations studied, and it would be interesting to find out if this is possible using the formalism developed in Chapter 7. However, as argued in Chapter 7 and (Broekens & DeGroot, 2006), emotion psychologist have to also formally annotate the data resulting from, and proposed mechanisms derived from emotion studies. Without this annotation, the computer model can not be evaluated other than in ways done in this thesis or in the work by many other modelers. So, formal modeling by computer scientists is only half the solution, and in this case, half a solution is no solution as there is nothing formal to compare the computer scientist’s formal model with. More importantly, the formalism proposed in Chapter 7 is targeted towards cognitive appraisal theory, which is not used as underlying theory for the research in Chapter 3 to 6. We have taken this direction because the number of computational models of emotion based on cognitive appraisal theory is vast, and consequently a formalism targeted at this family of models and theories could have a larger impact.

Second, we can definitely draw conclusions related to psychology from our studies, given that we extensively argued why we modeled affect in the ways we did, as well as how we used it to influence learning. Further, our conclusions

should be interpreted as mechanism existence proofs than can inspire psychological research, just as psychological research has inspired the modeling work in this thesis. Research can be done in many ways; sometimes the conclusions are clear-cut logical results, sometimes they are hypotheses made plausible. Our conclusions regarding computational results, such as better learning performance, fall into the first category: whatever the underlying mechanisms are or are not based upon, the result is objectively measurable. Conclusions related to the psychological implications of the studies presented in this thesis fall into the second category: given the computational results, the relations and mechanisms we have modeled become more plausible psychologically, although never an exclusive truth.