next up previous contents
Next: Discussion and Properties Up: Action-Reaction Learning: An Overview Previous: Action-Reaction Learning: An Overview

A Typical Scenario

Action-Reaction Learning (ARL) involves temporal analysis of a (usually multi-dimensional) data stream. Figure 2.3 displays such a stream (or time series). Let us assume that the stream is being generated by a vision algorithm which measures the openness of the mouth [43]. Two such algorithms are being run simultaneously on two different people. One person generates the dashed line and the other generates the solid line.

Figure 2.3: Dialog Interaction and Analysis Window

Now, imagine that these two individuals are engaged in a conversation. Let us also name them Mr. Solid (the fellow generating the solid line) and Mrs. Dash (the lady generating the dashed line). Initially (interval A-B on the time axis), Mr. Solid is talking while Mrs. Dash remains silent. He has an oscillatory mouth signal while she has a very low value on the openness of the mouth. Then, Mr. Solid says something shocking and pauses (B-C). Mrs. Dash then responds with a discrete 'oh, I see' (C-D). She too then pauses (D-E) and waits to see if Mr. Solid has more to say. He takes the initiative and continues to speak (E). However, Mr. Solid continues talking non-stop for just too long (E-G). So, Mrs. Dash feels the need to interrupt (F) with a counter-argument and simply starts talking. Mr. Solid notes that she has taken the floor and stops to hear her out.

What Action-Reaction Learning seeks to do is discover the coupling between the past interaction and the next immediate reaction of the participants. For example, the system may learn a model of the behaviour of Mrs. Dash so that it can predict and imitate her idiosyncrasies. The process begins by sliding a window over the temporal interaction as in Figure 2.3. The window looks at a small piece of the interaction and the immediate reaction of Mrs. Dash. This window over the time series forms the short term or iconic memory of the interaction and it is highlighted with a dark rectangular patch. The consequent reaction of Mrs. Dash and Mr. Solid is highlighted with the lighter and smaller rectangular strip. The first strip will be treated as an input ${\bf x}$ and the second strip will be the subsequent behavioural output of both Mr. Solid and Mrs. Dash (${\bf y}$). To predict and imitate what either Mr. Solid or Mrs. Dash will do next, a system system must estimate the future mouth parameters of both (stored in ${\bf y}$). As the windows slide across a training interaction between the humans, many such $({\bf
x,y})$ pairs are generated and presented as training data to the system. The task of the learning algorithm is to learn from these pairs and form a model relating ${\bf x}$ and ${\bf y}$. It can then generate a predicted ${\bf y}^*$ sequence whenever it observes a past ${\bf x}$ sequence. This allows it to compute and play out the future actions of one of the users (i.e. Mrs. Dash) when only the past interaction of the participants is visible.

Thus, the learning algorithm should discover some mouth openness behavioural properties. For example, Mrs. Dash usually remains quiet (closed mouth) while Mr. Solid is talking. However, after Solid has talked and then stopped briefly, Mrs. Dash should respond with some oscillatory signal. In addition, if Mr. Solid has been talking continuously for a significant amount of time, it is more likely that Mrs. Dash will interrupt assertively. A simple learning algorithm could be used to detect similar ${\bf x}$ data in another situation and then predict the appropriate ${\bf y}$ response that seems to agree with the system's past learning experiences.

Note now that we are dealing with a somewhat supervised learning system because the data has been split into input ${\bf x}$ and output ${\bf y}$. The system is given a target goal: to predict ${\bf y}$from ${\bf x}$. However, this process is done automatically without any manual data engineering. One only specifies a-priori a constant width for the sliding window that forms ${\bf x}$ and the width of the window of ${\bf y}$ (usually, the width will be 1 frame for ${\bf y}$to conservatively forecast only a small step into the future). The system then operates in an unsupervised manner as it slides these windows across the data stream. Essentially, the learning uncovers a mapping between past and future to later generate its best possible prediction.

next up previous contents
Next: Discussion and Properties Up: Action-Reaction Learning: An Overview Previous: Action-Reaction Learning: An Overview
Tony Jebara