This first simulation is based on three days of picking: day 1 training; days 2 and 3 for testing.
The results of the RL model are compared to two simple route plans
- Random: randomly select a route from the shortest distance, the shortest travel time route, and the congestion avoidance route
- Congestion: always chose the congestion avoidance route
It is surprising to see that the productivity reward works less than the speed reward approach. Maximizing the productivity of each AGV may not be the best approach to vehicle collaboration to ensure high global productivity.
The congestion strategy works well and requires less computational resources compared to the RL approach when congestion is the main bottleneck (i.e., when you have a large number of vehicles at the same time).
These results are based on a specific layout with only two days of picking. To gain a better understanding of this approach, in the following article, I will explain how to construct an AGV pick-up simulator and implement routing strategies.
This model should be tested with multiple subscription profiles in order to test the impact on productivity by tuning
- Number of rows per order (moves per order)
- Number of units collected per line
- Selection of active SKUs
The choice of strategy may vary if you have a campaign event in a particular SKU group, a shopping festival (Black Friday, 11/11) or out of season.