Attentional Agents and robot control

Abstract

A new approach based on attention control ideas is presented. We introduce the framework of Attentional Agents as the theoretical framework for constructing enhanced agents. We propose an approach where an agent processes multiple, possibly conflicting, goals in parallel and it uses a number of signals for obtaining information about its environment and own state. Additional requirements include a fast response in a changing environment and working with incomplete information. The proposed framework integrates in a seamless manner, sub-symbolic and symbolic processing, feedback control, conflict resolution, processing of multiple concurrent goals and action generation. We apply the general framework to the development of a robotic agent and we provide results from simple simulations where usage of multi-modal information and competing goals are present so as to illustrate the general ideas.

Combining Attention and Value Maps

“…The Attentional Agent architecture was thoroughly discussed in [4]. An Attentional Agent is a system which has the following major components: 1.…”

“…Formula (1) is an extension of the corresponding formula in [4]. ActInd is the action index of goal, which effectively controls the priority of the goal in global competition.…”

“…1, formula (1) for selecting goal priorities and we define the sensory and motor attention indices as in [4,5]. The boundary attention index is defined as a sigmoid function over the value map of the energy states of the agent, and it is given by (2):…”

“…The intrinsic weights in (1), which code the relative importance of goals are given as relative ratios: |Val G |/∑ |Val G | of the values for each goal, in our case: +20 for Transport (point B), +10 for Power Monitor (point C), -10 for MObj collision, -5 for SObj collision, +1 for Obj+ collision, -1 for Obj-collision and 5 for Collision Avoidance respectively. When approaching one of SObj, MObj, Obj+, Obj-closer than a threshold range R=3 cells, we calculate the S-AI and M-AI indices as described in [4,5]. Adding their contribution to (1) allows the alternation of priorities of goal execution and thus the determination of the currently active goal.…”

“…It has been modelled in a control theory framework by the second author [3]. Inspired by the above developments we have developed recently the Attentional Agent architecture [4] which combines a goalsbased computational model with an attention mechanism for selecting priority of goals dynamically in run time, for handling novelty and unexpected situations as well as learning of forward models based on the level of attention [5]. We now extend this model further to allow the attention mechanism to act as an alarm system when we approach limiting conditions.…”