An Ensemble of Cooperative Extended Kohonen Maps for Complex Robot Motion Tasks

Abstract

Self-organizing feature maps such as Extended Kohonen Maps (EKMs) have been very successful at learning sensorimotor control for mobile robot tasks. This paper presents a new ensemble approach called cooperative EKMs with indirect mapping to achieve complex robot motion. An indirect-mapping EKM self-organizes to map from the sensory input space to the motor control space indirectly via a control parameter space. Quantitative evaluation reveals that indirect mapping can provide finer, smoother, and more efficient motion control than does direct mapping by operating in a continuous, rather than discrete, motor control space. It is also shown to outperform basis function neural networks. Furthermore, training its control parameters with recursive least squares enables faster convergence and better performance compared to gradient descent. The cooperation and competition of multiple self-organized EKMs allow a non-holonomic mobile robot to negotiate unforeseen, concave, closely spaced, and dynamic obstacles. Qualitative and quantitative comparisons with neural network ensembles employing weighted sum reveal that our method can achieve more sophisticated 1 The final version of this article has been published in

Cooperative agent-based SANET architecture for personalised healthcare monitoring

“…Reinforcement learning is driven by continuous control space functions [17], as indirect-mapping methods provide fluid decision choices than direct mapping. The accuracy in sensory stimuli control is important where external factors directly affect the network's robustness and reliability, allows for developing a feature map approach using co-operative EKMs with indirect mapping to improve the responsiveness of the tracking process [4] [17].…”

“…The solution is to transform the views and decompose them into a singular map of the highest activation energies or stimuli, which is achieved once the maximum value of the integration matrix is calculated for all EKMs for each of the input stimuli received from the SANET's environment. Insufficient sensory controls can result in unexpected or undesirable outcomes, leading to possible faults in navigating a route to the gateway [17] [18]. When EKMs are established in the weighted-sum ensemble, a similar problem of unsolvable outcomes also takes place.…”

“…When EKMs are established in the weighted-sum ensemble, a similar problem of unsolvable outcomes also takes place. This can be solved by using an indirect EKM mapping method, whereby the control vector is manipulated indirectly via a control parameter space [17].…”

“…Using a multi-objective approach with co-operative EKMs as the modeling function has been adopted, with a reinforcement learning technique chosen on the basis of the situation or context [17] in an agent-based domain space. Interleaved EKMs that cooperate and compete to selforganize can enable a node to optimize in managing clustering and routing dynamically, whereby the node's output control is less than the total variable control available.…”

“…Utilizing an Extended Kohonen Map (EKM) by Kohonen [17] [18], the self-organizing map partitions sensory-aware spaces discretely; thus the generalization capability arises from its selforganization during map training [19]. Map resolution is improved in frequently encountered stimuli regions, thus mimicking biological sensory perceptions where reinforced practice allows for prediction of anticipated events.…”