Hardware-Assisted Large-Scale Neuroevolution for Multiagent Learning.

摘要

This DURIP equipment award was used to purchase, install, and bring on-line two Berkeley Emulation Engines (BEEs) and two mini-BEE machines to establish an FPGA-based high-performance multiagent training platform and its associated software. This acquisition of BEE4- W (Berkeley Emulation Engine) hardware platforms has costed $201,500.00 in total. To accelerate both the multiagent software simulation and hardware development, a DELL integrated cluster environment unit was purchased, featuring 192 2.8 GHz processors (on 24 nodes) with 384 GB high- speed RAM on an Infiniband backbone. We also installed cluster control software and licenses for the UBUNTU-based Linux operating system. Soft-ware was installed including JAVA and C/C++ compiler suites (debuggers, performance measurements, etc.), the Xilinx ISE software suite, and the Matlab software. The system was ordered in late 2012 and arrived in early 2013 and was up and running routinely with local users in a few weeks.