Skip to content

Commit

Permalink
added koos2011transferability
Browse files Browse the repository at this point in the history
Merge branch 'master' of yosinski.com:quadratot

Conflicts:
	paper_sean/results.tex
  • Loading branch information
suchanlee committed Jan 16, 2013
2 parents 4fe51ac + 0b7c113 commit a08063e
Show file tree
Hide file tree
Showing 2 changed files with 12 additions and 9 deletions.
2 changes: 1 addition & 1 deletion paper_sean/main.tex
Original file line number Diff line number Diff line change
Expand Up @@ -17,7 +17,7 @@
%\title{Evolving Quadruped Gaits for Physical Robots with a Bio-Inspired Generative Encoding Using a Simulator}
\title{Evolving gaits for physical robots with the HyperNEAT generative encoding: the benefits of simulation}

\author{Suchan Lee\inst{1} \and Jason Yosinski\inst{1} \and Kyrre Glette\inst{2} \and Hod Lipson\inst{1} \and Jeff Clune\inst{1}\inst{3} }
\author{Suchan Lee\inst{1} \and Jason Yosinski\inst{1} \and Kyrre Glette\inst{2} \and Hod Lipson\inst{1} \and Jeff Clune\inst{1}$^,$\inst{3} }
\institute{Cornell University, USA\and University of Oslo, Norway \and University of Wyoming, USA\\
\email{\{sl746,jy495,hod.lipson,jeffclune\}@cornell.edu, kyrrehg@ifi.uio.no}
}
Expand Down
19 changes: 11 additions & 8 deletions paper_sean/results.tex
Original file line number Diff line number Diff line change
Expand Up @@ -4,10 +4,14 @@
%to available data from previous studies.
%\subsubsection{Simulation:}

The simulator enabled HyperNEAT to evolve
fast, organic gaits. In simulation, HyperNEAT gaits were significantly faster than those from Glette et al. (\figref{threeplots}-Top, $\emph{p} < 6.8\times10^{-8}$, comparing the best gaits in the final generation of each run via MATLAB's one-sided Wilcoxon rank sum test).
The simulator enabled HyperNEAT to evolve fast, organic gaits. In
simulation, HyperNEAT gaits (\figref{threeplots}-Top) were
significantly faster than those from Glette et al, $\emph{p-value} <
6.8\times10^{-8}$ when comparing the best gaits in the final
generation of each run via Matlab's Wilcoxon rank sum test.
Specifically, HyperNEAT gaits were 52.1\% faster in simulation (25.4
cm/s, $\pm$ 3.4 SD versus 16.7 cm/s $\pm$ 1.9 SD, \tabref{results} -- qualitative conclusions remain the same even if the analysis is conducted using medians, but to make it easier for comparisons with earlier works~\cite{glette}, we use means $\pm$ SD).
cm/s $\pm$ 3.4 SD versus 16.7 cm/s $\pm$ 1.9 SD). Note that \tabref{results} reports
mean $\pm$ SD to make comparison with earlier work~\cite{glette} easier, but qualitative conclusions remain the same even when the analysis is conducted using medians.
Plots of servo positions over time reveal that the evolved HyperNEAT gaits are regular, smooth and coordinated~(\figref{threeplots}-Left), confirming previous results with HyperNEAT in simulation~\cite{clune2011performance,clune2009evolving}.

%\figOneOnTwo[h!]{threeplots}{med_fit_gen}{0.62}{servo_plot_111}{.48}{frequencyPlotThreshFinal}{.48}
Expand All @@ -16,8 +20,8 @@
{
\textbf{Top: }
%HyperNEAT outperforms a genetic algorithm with a simple encoding when both algorithms are combined with a simulator. Plotted are medians over 20 runs in simulation (solid lines) $\pm$ CIs (dashed lines). HyperNEAT gaits are 52.1\% faster in
HyperNEAT outperforms a genetic algorithm with a simple encoding when both algorithms are combined with a simulator. Plotted are means over 20 runs in simulation (solid lines) $\pm$ SD (dashed lines). HyperNEAT gaits are 52.1\% faster in
simulation and 5.1\% faster in reality~(\tabref{results}).
HyperNEAT outperforms a genetic algorithm with a simple encoding \cite{glette} when both algorithms are combined with a simulator. Plotted are means over 20 runs in simulation (solid lines) $\pm$ SD (dashed lines). HyperNEAT gaits are 52.1\% faster in
simulation and 5.1\% faster in reality than those from a previous study \cite{glette}~(details in \tabref{results}).
\textbf{Left: }
Servo positions over time (for nine servos) for a representative simulated HyperNEAT gait. HyperNEAT produced smooth and symmetrical gaits that contained complex regularities.
\textbf{Right: }Mean gait frequency averaged over 20 runs.
Expand Down Expand Up @@ -63,7 +67,7 @@
\hline
\end{tabular}
\vspace{.35cm}
\caption{Velocities of evolved gaits in simulation and on two different copies of the QuadraTot robot. Subject to availability, data are reported from the experiments for this paper and three previous studies. Reported are the total number of evaluations per run, the average of the fastest gaits produced in each run in simulation, and the single fastest gait produced on the CCML and ROBIN copies of the QuadraTot robot (see text for their differences). *Instead of using the single fitness value reported in~\cite{glette}, we ran 19 additional runs and used the average fitness of those 20 runs. Velocities are in cm/s, and bold indicates the best performance. **Median fitnesses of this paper: 26.9 cm/s with 95\% confidence intervals 23.8 cm/s and 26.75 cm/s.}
\caption{Velocities of evolved gaits in simulation and on two different copies of the QuadraTot robot. Subject to availability, data are reported from the experiments for this paper and three previous studies. Reported are the total number of evaluations per run, the average of the fastest gaits produced in each run in simulation, and the single fastest gait produced on the CCML and ROBIN copies of the QuadraTot robot (see text for their differences). *Instead of using the single fitness value reported in~\cite{glette}, we ran 19 additional runs and used the average fitness of those 20 runs. Velocities are in cm/s, and bold indicates the best performance. **Fitness median 26.9 cm/s, 95\% confidence interval [23.8 cm/s, 26.75 cm/s].}
\tablabel{results}
\end{center}
\end{table}
Expand Down Expand Up @@ -105,5 +109,4 @@
each servo, such as the Aracna platform~\cite{lohmann2012aracna}.
%which was designed in response to the challenges experienced with the
%QuadraTot platform.
That we did not model the servos in simulation, especially with their frequent failures, suggests that even better results could be obtained via a simulator that contained or learned servo models. In future work we will also incorporate techniques to minimize the gap between the simulator and reality~\cite{koos2010crossing,bongard,zagal}.

That we did not model the servos in simulation, especially with their frequent failures, suggests that even better results could be obtained via a simulator that contained or learned servo models. In future work we will also incorporate techniques to minimize the gap between the simulator and reality~\cite{koos2010crossing,bongard,zagal,koos2011transferability}.

0 comments on commit a08063e

Please sign in to comment.