Capítulo IV. El proceso de decisión en el consumidor
2. El reconocimiento de la necesidad
In this thesis, we presented a framework for human control of robots with low-fidelity input mechanisms, such as the mechanisms used in EEG-based brain-machine interfaces, which produce input commands that are noisy, discrete and that have high latency. We considered that although the input mechanism is low-fidelity, there is often a high-fidelity feedback mechanism that the robot can use to provide stimuli to the human user. Our objective was to enhance the performance of the interfaces for human control of robots by having humans provide input commands in response queries that are posed by the robot to maximize information gain rates for quick and reliable estimation of the human intent. We computed information gain rate of a particular query as the expected amount of information to be received per unit of time about the human intent by posing that query.
10.1.1 Discussion of the optimal feedback policy
Our first approach to querying human intent was to leverage an optimal feedback communication policy (Chapter 3), which was designed for reliable transmission of a message point between two computational agents. We presented the use of this policy in the context of human control of robots where the agent transmitting the message is a human user, and the message represents the human intent. This policy assumed a particular structure on the human intent, specifically that it can be mapped to a message point distributed uniformly in a closed real interval, and that there is an intuitive ordering between all possible human intents. We showed that such a structure exists when we represent human intent as a desired path to be followed by a
robot, based on the following two models. In the first model, desired paths were represented as strings of symbols chosen from an ordered alphabet, where each symbol identified a path primitive, such as a fixed-length circular arc with a particular curvature (Chapter 5). In the second model, desired paths were represented as local geodesics with respect to a cost function that took into account environment-driven features such as proximity to obstacles (Chapter 6). We demonstrated the feasibility of using the optimal feedback policy in robotic navigation tasks by querying the desired paths the humans want the robot to follow. In particular, we developed interfaces for flying a simulated aircraft and for navigating a mobile robot indoors with only noisy binary inputs obtained from EEG.
10.1.2 Discussion of the EEG-based interfaces for robotic
navigation
The interface for flying a simulated aircraft represented desired paths as strings of circular arcs, which corresponded to paths of piecewise-constant curvature, and allowed users to navigate the aircraft successfully over these strings at a fixed speed and altitude (Chapter 7). Experimental results with this interface showed that our approach outperformed an existing state-of- the-art approach in navigating a robot moving at a fixed speed and enabled successful specification of desired paths despite very low information transfer rates obtained with binary motor imagery decoding from EEG signals.
The interface for navigating a mobile robot in a virtual indoor environment represented desired paths as local geodesics with respect to a cost function that was recovered from human-demonstrated paths using structured predic- tion (Chapter 8). Effectively, this interface allowed a human user to navigate a robot by specifying the topological structure of their desired path, and letting the robot determine the geometric structure using the cost function learnt from existing data. In experiments, the robot was navigated along two target paths successfully in less than twice the time it would have taken to robot if informed of the path explicitly before the task.
Prior to these two interfaces, the existing interfaces for human control of robots with EEG heuristically used inputs commands, which were inherently noisy and low-bandwidth, to identify moment-to-moment steering commands
(i.e., in the interfaces based on process control such as [43]), waypoints (i.e., in the interface by Iturrate et al. [53]) or final destinations (i.e., in the interfaces based on goal selection such as [51]). In contrast, our interfaces used input commands as evidence to infer the user’s desired path over a compact space of desired paths. The experimental results showed that our approach is not only optimal for reliable communication of intent but also generates a protocol that is easy for humans to implement for navigating robots.
10.1.3 Discussion of the active inference policy
Our second approach to querying human intent was to leverage the Bayesian active inference framework (Chapter 4). In particular, we presented the use of a (sequential) Bayesian experimental design, which suggested the selection of the experiment with the maximum information gain about the unknown parameter to be learnt, in the context of designing queries for inferring hu- man intent, where each query specified the input command the human must provide to communicate information about their intent. To evaluate the value of a query, we introduced a measure we called information gain rate (IGR), which was defined as the ratio of the information gain to the expected latency of observing an input command after a query was posed. However, other measures could be used to evaluate the value of a query depending on the objectives of the interface design.
We also established a link between the active inference policy that max- imizes information gain rates and the optimal feedback policy presented in Chapter 3. In particular, the queries selected by these two policies are the same if the human intent can be mapped to a particular structure (i.e., to a message point uniformly distributed over a closed real interval), the human responds to each query using the same input mechanism, and the expected time to obtain an input command after posing a query is the same for all queries. If these conditions are satisfied, the active inference policy reduces to the optimal feedback policy.
10.1.4 Discussion of the EEG-based interface for text-entry
We demonstrated the feasibility of using the active inference policy by de- veloping an EEG-based interface for text entry (Chapter 9). Users of this interface specified desired texts by responding to a sequence of queries that were chosen adaptively to maximize information gain rates. To our knowl- edge, our interface was the first to design queries to make the input commands received from human users have the maximum value for identifying the user’s desired character based on the characteristics of the input mechanism. The experimental results showed that our interface allowed human users to specify text twice as fast as they would with the compared state-of-the-art interfaces using the same input mechanism.