...
 
Commits (2)
......@@ -613,7 +613,4 @@ In the third experiment, we simplified the problem by breaking it apart into sub
It is quite evident that the constraints of having to generate a capable controller for targeted locomotion in a reasonable time to consider a hardware implementation, are not suitable for a black box approach.
Learning movement logic and target logic as a whole with the same algorithm is a problem more suitable for environments where long learning times are feasible.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\subsubsection{Hardware testing}
Hardware testing and verification were never performed because the results in simulations are not satisfying.
......@@ -52,10 +52,20 @@ The task is too complicated for our controller to learn all at once
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
\section{Considerations about the approach}
The amount of time required to collect all data was 25 days of continuous computing, considering all the different configurations presented in this thesis.
We divided the work between two machines with respectively 8 cores and 4 cores.
The total CPU time estimations are measured to be around 264 days, roughly 9 months. The total amount of simulated time is estimated to be 23 months.
These time measurements exclude several other tests, not reported in this thesis, that were conducted with different configurations, mainly preliminary tests and extended runs.
We did not achieve our research goal and we have to have reflected on what went wrong and what could have been done to improve the results.
The collection of the required data for the gait part of the experiment took a long time, 16 days in total (192 days of CPU time, 500 days of simulated time), as it did not include a racing mechanism.
We ran the phototaxis part of the experiment with the racing mechanism activated, reducing the average CPU time spent on each evaluation from 11s to 3s.
The total time spent for the phototaxis part of the experiment was 3 days for each configuration (plain, kickstarting and separated lights) for a total of 9 days (72 days of CPU time, 189 days of simulated time).
The gait experiments were run on both the machines, the phototaxis experiments were run only on the 8 core machine.
We conclude that a black-box approach for such a complex task was a leap too big for our algorithm to make.
Sadly, we spent a lot of time running simulations but we did not achieve our research goal.
We have to reflect on what went wrong and what more could have been done to improve the results.
Our conclusion is that a black-box approach for such a complex task is a leap too big for our algorithm to make.
Evidence that the complexity of the problem was our main obstacle can be found in the results: every time there is a decrease in complexity, there is a corresponding boost in the performance of the results.
E.g. when we introduced the kickstarting hack, only the simple shapes (fewer control joints) could take benefit;
when we separated the light targets as different tasks, we found a decisive increase in all the results, with some more than doubled.
......