Genetic Algorithm Based Deep Learning Parameters Tuning for Robot Object Recognition and Grasping

Transcription

1 Genetc Algorthm Based Deep Learnng Parameters Tunng for Robot Obect Recognton and Graspng Delowar Hossan, Genc Cap Internatonal Scence Index, Mechancal and Materals Engneerng waset.org/publcaton/ Abstract Ths paper concerns wth the problem of deep learnng parameters tunng usng a genetc algorthm (GA) n order to mprove the performance of deep learnng (DL) method. We present a GA based DL method for robot obect recognton and graspng. GA s used to optmze the DL parameters n learnng procedure n term of the ftness functon that s good enough. After fnshng the evoluton process, we receve the optmal number of DL parameters. To evaluate the performance of our method, we consder the obect recognton and robot graspng tasks. Expermental results show that our method s effcent for robot obect recognton and graspng. Keywords Deep learnng, genetc algorthm, obect recognton, robot graspng. I. INTRODUCTION L s a vastly growng research feld. It tres to mmc the Dhuman bran. For example, the feature herarches n the human vsual cortex represent the obects at the dfferent level of abstracton. The more abstract features go up n the herarchy; the obects become more vsble to the human. DL works n the same way as our human bran organzes deas n herarchcal fashon. The obectve of DL s to brng the machne learnng research to Artfcal Intellgence. DL research was started by Hnton et al. [1] n After then, many researchers are workng to mprove the performance of DL. DL has many parameters whch have nfluence on performance. For ths reason, recently researchers are workng to ntegrate evolutonary programmng wth DL. Some research works are already presented. The frst effort was done by Lamos-Sweeney [2]. He ntegrated GA wth a multlayer DL network for data compresson and obect classfcaton. He showed that hs proposed method enhanced the flexblty and reduced the computatonal burden of the algorthm. Levy et al. [3] proposed a hybrd approach to ntegratng GA and deep restrcted Boltzmann machnes (RBMs) for panter classfcaton problem. They extracted features usng generc mage processng functon and deep RBMs. Trumala [4] studed the evolutonary computaton (EC) mplantaton possblty wth the deep archtectures. He argued EC could solve the Deep Neural Networks (DBNs) overfttng problem. Davd and Greental [5] proposed a GA-asssted method for a deep autoencoder to mprove the performance and produced a sparser neural network. Verbancscs and Harguess Delowar Hossan s wth the Graduate School of Scence and Engneerng for Educaton, Unversty of Toyama, 3190 Gofuku, Toyama , Japan (e-mal: delowar.hossan.38@hose.ac.p). Genc Cap s wth the Department of Mechancal Engneerng, Hose Unversty, Tokyo , Japan (correspondng author, phone: , fax: , e-mal: cap@hose.ac.p). [6] nvestgated a neuro-evoluton (NE) based DL method. They appled the Hypercube-based NeuroEvoluton of Augmentng Topologes (HyperNEAT) to tranng a feature extractor process n backward propagaton learnng. Shao et al. [7] developed an evolutonary learnng methodology based on multobectve genetc programmng (MOGP) for mage classfcaton. It generated doman-adaptve global feature descrptor automatcally. DL has demonstrated state-of-art performance on robotc applcatons. There have been many works presented for robot real-tme obect recognton and graspng [8]-[13]. The goal s that robots can learn through the nteracton wth the envronment and the human subects. It reduces the stress of human subects n varous home style and ndustral tasks. In ths paper, we present an evolutonary learnng method that combned GA and Deep Belef Neural Networks (DBNNs) for robot obect recognton and graspng. Ths method optmzes the DBNN parameters, such as the number of hdden unts, the number of epochs, learnng rates and momentum n learnng procedures of the hdden layers. It reduces the error rate and network tranng tme of obect recognton. The obects are recognzed n the dfferent orentaton, postons, and lghtng condtons usng optmzed DBNN method, after then the robot pcks up the recognzed obects and places n a predefned poston. Ths paper s organzed as follows. In Secton II, combned GA and DBNN are descrbed. In Secton III, GA parameters and results are presented. In Secton IV, the expermental results of optmzed DBNN for robot obect recognton and graspng are shown. In Secton V, we conclude and menton the future works. II. COMBINED GA WITH DEEP BELIEF NEURAL NETWORK (GADBNN) The problem wth DBNN s that the DBNN parameters need to set before the tranng begns. Although there are no any fxed rules how to set these parameters, these parameters have nfluenced on the success of the tranng. By combnng GA wth Deep Belef Neural Network (GADBNN), the DBNN parameters are optmzed usng GA. The flowchart of GADBNN s shown n Fg. 1. A. Structure of Deep Belef Neural Network (DBNN) A DBNN s a probablstc generatve model, whch s constructed by a stack of RBMs. An RBM conssts of a vsble layer and a hdden layer, or a hdden layer and another hdden layer. In RBM, the neurons of each layer are completely connected wth the neurons of another layer, but the neurons of 629

2 Internatonal Scence Index, Mechancal and Materals Engneerng waset.org/publcaton/ the same layer are not nternally connected wth each other. RBMs are stacked on the top of each other to buld a DBNN. DBNN extracts feature n the herarchcal fashon, where lower level features form hgher levels features. The energy functon between vsble layer and hdden layer { v, h} n RBM s gven as follows: where nv nh n n 1 1 v 1 h E( v, h) a b w a v b h (1) n s the number of vsble unts, v 1 n s the number of h hdden unts, a s the bas term for vsble unts, b s the bas term for hdden unts, w s the weghts between vsble and hdden unts, v s the vsble unts wth v {0,1}, and h s the hdden unts wth h {0,1 }. The structure of our DBNN s shown n Fg. 2. It conssts of a vsble layer, three hdden layers, and an output layer. The optmal number of hdden unts n the three layers was found by applyng GA [14]. Vsble unts are set to the actvaton probabltes; on the other hand, hdden unts are set to the bnary values. We use two types of samplng method, such as Contrastve Dvergence (CD) and Persstent Contrastve Dvergence (PCD). The frst hdden layer s sampled by PCD because PCD explores the entre searchng doman for the nput features. It s much better n representng the log-lkelhood of the pxels. The second and thrd hdden layers are sampled usng CD because CD explores the better near the nput mages. CD s not aware of spurous modes of the nput mages and better for extractng features. For ths reason, we combne CD and PCD samplng methods and t outperforms for obect recognton purpose. Fg. 1 Flowchart of combned GA and DBNN Backpropagaton error dervatves are used to reduce the dscrepances between the orgnal features and ts reconstructon for fne-tunng weghts n order to better obect recognton. It refers the whole procedure encompassng the calculaton of the gradent and uses n stochastc gradent descent. We use softmax functon for classfyng obects. Backpropagaton wll termnate, f t satsfes one of the followng condtons: (1) reach to best performance,.e. mean square error (MSE), (2) reach to maxmum epoch number,.e. 200, (3) reach to mnmum gradent value, or (4) reach to maxmum valdaton checks,.e. 6. Fg. 2 Structure of our DBNN A. Genetc Algorthm (GA) GA s the heurstc search algorthm based on the natural evolutonary deas. It s the best way to solve a problem that has lttle nformaton to know. It s a very general algorthm and sutable for any search space. GA takes advantages of hstorcal nformaton for searchng of better soluton among the search space. The basc technque s to desgn an evoluton process to smulate the survval of the fttest among ndvduals for solvng a problem n generatons. The evoluton process of DBNN s shown n Fg. 1. In GADBNN, the GA s used to fnd the optmal number of DBNN parameters, such as the number of hdden unts, the number of epochs, learnng rates and momentum n learnng procedures for hdden layers. DBNN nformaton s encoded n the genome of the GA. Intally, the random number of ndvduals s generated. Then, the number of DBNN parameters s evaluated and ranked. After then, the ftness functon s evaluated. If the convergence crtera are not satsfed, then crossover and mutaton creates new ndvduals and replaces the worst members of populaton. If the convergence crtera are satsfed, then evoluton process s termnated, and the optmal number of DBNN parameters s generated. III. GA PARAMETERS In ths paper, we are deployed a real-valued GA [14], whch generates better solutons wth respect to the qualty of the soluton. It uses real values as parameters of the chromosome n populatons. It s mplemented wth the selecton, crossover and 630

3 Internatonal Scence Index, Mechancal and Materals Engneerng waset.org/publcaton/ mutaton operators. The GA functon and parameters are mentoned n Table I. Populaton s dvded nto four subpopulatons, and each subpopulaton uses dfferent mutaton rates. Mgraton s selected based on the number and sze of the subpopulatons. Crossover probablty s same for all subpopulatons. The total number of ndvduals s 100, and the maxmum number of generaton s 30. The best subpopulaton has receved the resources, and worst ndvduals are removed from less successful generatons. TABLE I GA FUNCTIONS AND PARAMETERS Functon Name Parameters Number of subpopulatons 4 Intal number of ndvduals (subpopulaton) 25, 25, 25, 25 Crossover probablty 0.8 Mutaton rate (subpopulaton) 0.1, 0.03, 0.01, Isolaton tme 10 generatons Mgraton rate 10% Results on screen Every 1 generaton Competton rate 10% Termnaton 30 generatons The ftness functon s defned as to mnmze the error rate and mnmze the network tranng tme n order to optmze the number of hdden unts, the number of epochs, learnng rates and momentum n hdden layers. The ftness functon used n our mplementaton s shown as follows: Ftness ( T T DBP ) 100 ( E E ABP ) (2) 40 where E s the total number of msclassfcaton dvded by the total number of test data before backpropagaton, E ABP s the total number of msclassfcaton dvded by total number of test data after backpropagaton, T s the network tranng tme n second before fne-tuned operaton usng backpropagaton, T s the network tranng tme n second DBP durng fne-tuned operaton usng backpropagaton operaton. Fg. 3 Best obectve values per subpopulaton IV. EXPERIMENTAL RESULTS In ths secton, we descrbe the experment and results of our approach. We dvde the expermental results nto two sectons: (1) Expermental result for GA based Deep Belef Neural Network (GADBNN), (2) Expermental result for Optmzed DBNN for robot obect recognton and graspng. A. Expermental Results for GADBNN In order to optmze DBNN parameters, we buld a database, whch conssts of 1200 tran mages of sx dfferent types of obects (200 mages of each obect) and 600 test mages of 100 mages of each obect. Ths tranng and test mages were taken randomly n dfferent orental, postons, and lghtng condtons n our expermental envronment to make our system robust. In our dataset, all mages are grayscale mages. Fg. 4 Obectve values of all ndvduals of generatons In order to evaluate GADBNN performance, we plot the best obectve value per subpopulaton n Fg. 3. Four subpopulatons (sp1, sp2, sp3, sp4 where sps are the subpopulatons) are mentoned wth four dfferent colors. The best obect value s mentoned by bold red color. In addton, obectve values of ndvduals through all generaton durng evoluton process are shown n Fg. 4. The ftness value was started from In frst four generatons, the strategy appled n subpopulaton 2 s successful. On the ffth generaton, strategy 4 s successful. From sx to nne generatons, strategy 1 s successful. The ndvdual s convergence s most successful n the tenth generaton and obectve value s At the end of the optmzaton (tenth generaton), the strategy appled on subpopulaton 4 s the most successful. The ndvdual s convergences were termnated on 24th generaton. B. Expermental Results for Obect Recognton and Robot Graspng In order to demonstrate the performance of optmzed DBNN parameters, we consder the obect recognton and robot graspng tasks. For ths purpose, we buld an mage dataset that s conssted of 1200 mages (200 mages of each obect) of sx 631

4 World Academy of Scence, Engneerng and Technology Internatonal Scence Index, Mechancal and Materals Engneerng waset.org/publcaton/ robot graspable obects n dfferent orentatons, postons, and lghtng condtons n our expermental envronment. 1. Obect Recognton Results When a user requests for an obect by clckng on Graphcal User Interface (GUI), then a snapshot s taken by USB camera (shown n Fg. 6). Ths snapshot s converted to the grayscale mage. We apply a morphologcal structurng element operaton, and all exstng obects n the envronment are extracted and separated based on the center of the obects of the sze of 28x28 pxels. In the vsble layer, 784 neurons are used as nput. From optmzed DBNN, we found 650, 413, and 555 neurons of hdden unts, 100, 88, and 184 neurons of epochs n three hdden layers. Learnng rate for frst hdden layer was and for second and thrd hdden layers. Momentum values of frst fve epochs n each hdden layer are , , , , and n order to make learnng procedures more effcent. Momentum value for the remanng epochs n each hdden layer s 0. After passng through the three hdden layers, DBNN generates sx probablty values as output, because we traned optmzed DBNN wth sx dfferent types of obect. Each probablty belongs to each obect. The hghest probablty value s consdered the recognzed obect. If the hghest probablty value s smaller than the threshold value (0.7), then we consder that the requested obect does not exst n the expermental envronment. For example, the blue screwdrver s used as nput of DBNN, as outputs, DBNN generates sx probablty values, such as , , , , , and (shown n Fg. 5). From ths result, the maxmum probablty s , whch belongs to the fourth obect that we defned as the blue screwdrver. At the same way, other obects can be recognzed usng optmzed DBNN method. Fg. 5 Obect recognton process Fg. 6 Snapshots of robot obect recognton and graspng usng optmzed DBNN 632

5 Internatonal Scence Index, Mechancal and Materals Engneerng waset.org/publcaton/ Robot Graspng Results For robot graspng purpose, we used a Programmable Unversal Machne for Assembly (PUMA) robot from Denso Corporaton. It has a grpper that can grasp any obects of the maxmum wdth of 95 mm. When a user requests for a specfc obect, then the requested obect s recognzed usng optmzed DBNN method. The robot fnds the graspng poston based on the recognzed obect. The robot generates a moton traectory from the ntal poston to the obect graspng poston. After reachng to the obect graspng poston, the robot adusts ts grpper orentaton based on the obect orentaton n order to grasp obects n dfferent orentatons. After graspng the obect, the robot generates another moton from obect poston to the predefned obect placng poston. After placng the obect, the robot returns to the ntal poston and wats for next requests from the user. The snapshots of robot obect recognton and graspng process are shown n Fg. 6. V. CONCLUSIONS AND FUTURE WORKS The goal of ths paper s to propose a method to optmze DL parameters usng evolutonary algorthms. Our proposed method optmzed the DL parameters, such as the number of hdden unts, the number of epochs, learnng rates and momentum of learnng procedures n hdden layers. The error rates and network tranng tme were mnmzed. We evaluated our optmzed method on real-tme obect recognton and robot graspng process. The results showed that our optmzed method outperformed at the assgn tasks. In future, we wll ntegrate Mult-obectve Evolutonary Algorthms (MOEA) because the coeffcents of sngle obect evolutonary algorthms n the obectve functon are very dffcult to determne. In addton, we want to consder the scalng factors of the obects. REFERENCES [1] G. Hnton, S. Osndero, and Teh, Y.-W, A fast learnng algorthm for deep belef nets, Neural Computaton, vol. 18, no. 7, pp , [2] J. Lamos-Sweeney, Deep learnng usng genetc algorthms, Master s Thess, Rochester Insttute of Technology, NY, USA, [3] E. Levy, O. E. Davd, and N. S. Netanyahu, Genetc algorthms and deep learnng for automatc panter classfcaton, Proceedngs of the 2014 Annual Conference on Genetc and Evolutonary Computaton, 2014, pp [4] S. S. Trumala, "Implementaton of Evolutonary Algorthms for Deep Archtectures", AIC, pp [5] O. E. Davd, and I. Greental, "Genetc algorthms for evolvng deep neural networks", In Proceedngs of the Companon Publcaton of the 2014 Annual Conference on Genetc and Evolutonary Computaton, pp ACM, [6] P. Verbancscs, and Josh Harguess, "Generatve neuroevoluton for deep learnng", arxv preprnt arxv: (2013). [7] L. Shao, L. Lu, and X. L, Feature learnng for mage classfcaton va multobectve genetc programmng, IEEE Transactons on Neural Networks and Learnng Systems, vol. 25, no. 7, pp , [8] I. Lenz, H. Lee, and A. Saxena, Deep Learnng for Detectng Robotc Grasps, Internatonal Journal of Robotcs Research, vol. 34, no. 4-5, pp , [9] L. Pnto, and Gupta, Superszng Self-supervson: Learnng to Grasp from 50K Tres and 700 Robot Hours, arxv: (cs.lg), [10] D. Hossan, and G. Cap, Applcaton of Deep Belef Neural Network for Robot Obect Recognton and Graspng, The 2nd IEEJ Internatonal Workshop on Sensng, Actuaton, and Moton Control (SAMCON 2016), Tokyo, Japan, March [11] D. Hossan, G. Cap, and M. Jnda, Obect Recognton and Robot Graspng: A Deep Learnng based Approach, The 34th Annual Conference of the Robotcs Socety of Japan (RSJ 2016), Yamagata, Japan, September [12] J. Redmon, and A. Angelova, Real-Tme Grasp Detecton Usng Convolutonal Neural Networks, arxv: , [13] S. Levne, P. Pastor, A. Krzhevsky, and D. Qullen, D. Learnng Hand-Eye Coordnaton for Robotc Graspng wth Deep Learnng and Large-Scale Data Collecton, arxv: , [14] G. Cap, and K. Doya, Evoluton of recurrent neural controllers usng an extended parallel genetc algorthm, Robotcs and Autonomous System, vol. 52, no. 2-3, pp , Delowar Hossan s workng toward the Ph.D. degree at Graduate School of Scence and Engneerng for Educaton, Unversty of Toyama, Japan, also workng as vstng researcher at Hose Unversty, Japan. He receved the B.Sc. and M.Sc. degrees n Computer Scence & Engneerng from Unversty of Rashah, Rashah, Bangladesh, n 2010 and 2012, respectvely. He was a Lecturer at the Department of Computer Scence and Engneerng, Dhaka Internatonal Unversty, Dhaka, Bangladesh from 2012 to Hs research nterests nclude the ndustral robot, deep learnng, artfcal ntellgence, computer vson, mage processng, ntellgent robotcs, learnng, and evoluton. Genc Cap receved the B.E. degree from Polytechnc Unversty of Trana, n 1993 and the Ph.D. degree from Yamagata Unversty, n He was a Researcher at the Department of Computatonal Neurobology, ATR Insttute from 2002 to In 2004, he oned the Department of System Management, Fukuoka Insttute of Technology, as an Assstant Professor, and n 2006, he was promoted to Assocate Professor. In 2010, he was oned as a Professor at the Department of Electrcal and Electronc Systems Engneerng, Unversty of Toyama, Toyama, Japan. He s currently a Professor at the Department of Mechancal Engneerng, Hose Unversty, Tokyo, Japan. Hs research nterests nclude ntellgent robots, BMI, mult-robot systems, humanod robots, learnng, and evoluton. 633