Real-Time GA-Based Probabilistic Programming in Application to Robot Control
Possibility to solve the problem of planning and plan recovery for robots using probabilistic programming with optimization queries, which is being developed as a framework for AGI and cognitive architectures, is considered. Planning can be done directly
- PDF / 2,687,870 Bytes
- 11 Pages / 439.37 x 666.142 pts Page_size
- 63 Downloads / 169 Views
ITMO University, St. Petersburg, Russia St. Petersburg State University, St. Petersburg, Russia 3 AIDEUS, St. Petersburg, Russia [email protected], [email protected], [email protected] Aix Marseille Université, CNRS, LAM (Laboratoire d’Astrophysique de Marseille) UMR 7326, 13388 Marseille, France 2
4
Abstract. Possibility to solve the problem of planning and plan recovery for robots using probabilistic programming with optimization queries, which is being developed as a framework for AGI and cognitive architectures, is considered. Planning can be done directly by introducing a generative model for plans and optimizing an objective function calculated via plan simulation. Plan recovery is achieved almost without modifying optimization queries. These queries are simply executed in parallel with plan execution by a robot meaning that they continuously optimize dynamically varying objective functions tracking their optima. Experiments with the NAO robot showed that replanning can be naturally done within this approach without developing special plan recovery methods. Keywords: Probabilistic programming Optimization queries algorithms Robot planning Replanning
Genetic
1 Introduction It is frequently assumed that AGI systems should not only perform some abstract reasoning, but should also be able to control some body achieving goals in real environments. Even if a cognitive architecture wasn’t initially developed specifically for this purpose, natural desire to try applying it for e.g. robot control can arise after its maturing. Robot control tasks are quite interesting since they require both planning and reactive control for achieving a goal in dynamic environments. Symbolic architectures are usually good for planning, but realization of reactive behavior within them is awkward, while emergent architectures are usually better suited for reactive control. Thus, hybrid solutions are developed to solve the problem of plan recovery [1]. We are developing an approach to AGI using probabilistic programming as the starting point. In another paper [2], we explain motivation behind this approach and discuss how traditional probabilistic programming languages (PPLs) should be
© Springer International Publishing Switzerland 2016 B. Steunebrink et al. (Eds.): AGI 2016, LNAI 9782, pp. 95–105, 2016. DOI: 10.1007/978-3-319-41649-6_10
96
A. Potapov et al.
extended in order to become usable as a framework for development of cognitive architectures. However, this discussion addresses questions regarding reasoning and learning, but not regarding controlling (embodied) agents. At the same time, most general-purpose PPLs support only computationally expensive queries with unpredictable execution time (so they are well-suited for planning, but not for reactive control). This issue might be called purely technical, but nevertheless it is quite important. Indeed, taking limitation of resources into account is considered essential for AGI research [3]. In this paper, we describe how PPLs with optimization queries based on
Data Loading...
