AWS India have announced the DeepRacer Women’s League to promote machine learning among the students in India. Racers had pretty much two weeks to learn how to train their models and to qualify into the final. Let’s see how they are doing.
AWS DeepRacer
AWS DeepRacer is a 1/18th scale autonomous race car but also much more. It is a complete program that has helped thousands of employees in numerous organizations begin their educational journey into machine learning through fun and rivalry.
Visit AWS DeepRacer page to learn more about how it can help you and your organization begin and progress the journey towards machine learning.
Join the AWS Machine Learning Community to talk to people who have used DeepRacer in their learning experience.
AWS DeepRacer Women’s League India
AWS India have organised an event with an intention to promote diversity in tech. They have invited students from across the country to take part in a program which included a few sessions:
After the training racers first competed in four regional races to qualify for the National Race from which 15 best racers qualified to the last stage. Then they had an extra training with AWS experts to help them prepare for the Grand Finale which took place on the 21st of April. Best ones would get pretty sweet prizes:
Great Finale
The format of the race is as follows:
- The race format is time trial
- The track is European Seaside Circuit (60 meters long, 1.28 meters wide, very demanding)
- Each racer has one 3 minutes time slot to clock a lap as good as possible
- Each time the car leaves the track it gets reset and a 3 second penalty is applied
- During the race a racer is given the ability to modify the car’s speed values (cars usually have a list of actions that they choose from). This comes in form of a slider which has a multiplier value applied to all actions. This gives an opportunity and a risk – cars need to know how to operate in a given speed, also some actions might get broken at higher speeds
You can watch the whole final here:
… or read our summary below.
First to race was Shreya Sajid of Loyola-ICAM College of Engineering and Technology. She started training relying on waypoints and then moved one to a strategy involving number of steps and progress. She started with a lap of 43.985 and then improved to 42.283 and 42.147. The last lap has been affected by multiple resets and with 41.147 Shreya finished her round.
Anupriya Shree of IIT Dhanbad managed to go into lead with her first lap of 40.822 and then managed to improve to 38.485, then 37.404 and eventually 34.931. Anupriya has started training using the centre line strategy and then progressed to a strategy that involved optimising the racing line for the car. Also she played around with hyperparameters: she lowered the discount factor from 0.999 to 0.99 to promote more local optimization of behaviour and reduced the entropy gradually as she cloned the models to stimulate learning towards a local optimum. In longer races this might be a problem as the car could stop looking for an even better strategy, but with so limited time it helped the model focus on a certain inference pattern and squeeze as much as possible out of it. 34.931 has put Anupriya in a strong lead.
Pearl Patel of Vidyalankar Institute of Technology has started preparing for the training by playing video games and observing how racing cars behave on tracks. Based on this experience and her intuition. From the looks of it Pearl has optimised for behaviour on the turns – she admitted that she looked for the optimal line on the track and trained the car to follow that. She also added a reward based on the amount of steps on completion. Unfortunatel her model lacked stability to squeeze more out of it and thus she ended her run with 38.213s and a few more laps in the area of 41.4-44.5s.
Mehak G. of India Institute of Technology has struggled with the stability and with the resets that she experienced she clocked laps of 41.189 and 39.674. As shallowracer1 said on Twitch chat, stability is the X factor in the race.
Himani Ramwani of LJIET focused on rewarding based on progress, speed and steering. Her model was pretty stable but pretty decent times of 39.506 and more in the area of 41-45 seconds weren’t fast enough to get to the top. 3 seconds reset does not give enough incentive for playing it safe.
Sawi Sharma of Panjab University Swami Sarvanand Giri (SSG) Regional Centre has had a really strong start – first lap of 34.769 seconds has put her in first place straight away! She has trained her model for about six hours and mainly focused on following the centre line. Seems like her first lap was on the edge of the car’s performance – she pushed her speed a little bit getting into stability problems. Still, her laps were in the 35-37s area which is very impressive.
Helen Thomas of Indian Institute of Science Bangalore has done some insane racing there. Her model was pretty stable and fast. She started with 34.969 which would put her in third but she quickly improved with 33.330 taking the top spot for now. She probably pushed the speed a bit too much as she started then struggling with the resets. Still she was going in the 33-34 second region (if you ignore the resets).
Nandini Malviya of Bharati Vidyapeeth’s College Of Engineering For Women in Pune followed. Seems the model was mainly going along the centre line and trying to cut the corners a little bit. There s a rather big problem in the mild turns section of the track – it’s pretty difficult to get the car to go stably there so quite often the car just goes crazy there. Her best lap of 43.223 is pretty good but not enough to compete for the highest spots.
Kashish Bansal of IIT Indore started pretty fast. Her opening lap of 44.961 with two resets was pretty good. Second lap of 40.337s with a single reset means she started getting some noticeable improvements. Kashish switched to higher speed now and with no resets she improved over the whole track and with 33.728 she jumped into second. Fourth lap struggled with resets and with 33.728s
Sugandha Tanwar of Dayal Bagh Educational Institute started with centre line reward and then moved through progress and steps (which usually are a rather time consuming method of training) and then moved on to training with use of waypoints and the race line, and tracking progress with log analysis. She started with a pretty decent 40.186s time and then improved to 39.739 and then 37.645 with a single reset. Seems that this was the edge of the model’s stability as she started experiencing many more resets.
Surbhi Agrawal of Visvesvaraya National Institute of Technology started very strong with the best time in sector one across all the racers. Good enough to clock 33.380s despite a reset! In lap two she faced a reset in sector one which meant that despite clocking the best time in sector two she couldn’t improve. Third and fourth laps were a replay of the second – reset in sector one. If it wasn’t for the reset each of those laps would have put her in the top spot. We’re reaching the magical point where stability becomes the X factor next to speed. So close! But not close enough, 33.380s put her in the second spot.
Jayita Pramanik of Guru Nanak Institute of Technology started rather safe with the speed. First lap of 42.358 followed a nice line but just wasn’t fast enough. In second lap she improved across all the seconds and hit 37.530 s which is pretty good. Jayita trained based on the waypoints and learned from the AWS DeepRacer resources that AWS have shared. Thrid lap she improved significantly in second sector and clocked 35.401, very nice!
Sakshi Maheshwari of the Institute of Engineering and Technology, DAAV, Indore started rather safe with just over 40 and then 39.466 and again 44 s after a few resets. She definitely had the potential for improvement if it wasn’t for the resets. She managed to shave a little bit with the last lap of 39.144 which has left her hungry for a better result. It’s a pity, I’m sure there was more in the model but the stability just denied it.
Varsha Rani of Rajasthan Institute of Engineering and Technology focused on rewarding the racing line and general behaviour in terms of speed. She experimented with many models and chose the one she believed was best. First lap of 47.321 with three resets confirmed one of my worries – when you reward just going fast you can reward the car for going off track fast if you don’t balance it with other components of your reward function. Second lap above 50 seconds with 4 resets just confirms it. This track is really unforgiving to that. In third Varsha slowed down a little bit and thanks to it she managed to clock 43.862 s thanks to the stability gains. She was surely not satisfied with the outcome but she should be proud of what she’s learned. The results are a mixture of skill and luck and the skill is definitely there. Well done!
Swasti Khurana of Indian Institute of Information Technology Vadodara started pretty fast with the first lap of 38.521 with one reset so there was a potential for improvement. Next one with 38.064s also with a single reset just confirmed that there is a potential. In third she had one reset again but was able to clock 33.800 s just by improving in sectors one and three. Then she sadly experienced a few more resets and a bit of slowdown but this is a great time.
Congratulations to Helen, Surbhi and Kashish for making it to the podium:
Let’s have a look at the end results:
| Racer | Time | |
| 1. | Helen Thomas | 33.330 |
| 2. | Surbhi Agrawal | 33.380 |
| 3. | Kashish Bansal | 33.728 |
| 4. | Swasti Khurana | 33.800 |
| 5. | Sawi Sharma | 34.769 |
| 6. | Anupriya Shee | 34.931 |
| 7. | Jayita Pramanik | 35.401 |
| 8. | Sugandha Tanwar | 37.645 |
| 9. | Pearl Patel | 38.213 |
| 10. | Sakshi Maheshwari | 39.144 |
| 11. | Mehak G. | 39.674 |
| 12. | Himani Ramwani | 39.506 |
| 13. | Shreya Sajid | 42.147 |
| 14. | Nandini Malviya | 43.223 |
| 15. | Varsha Rani | 43.862 |
What I find really impressive is that these racers have mostly not done any reinforcement racing before and started playing and learning with DeepRacer two weeks ago. Sure, some of their methods worked out better than others, also there is a bit of randomness since we are applying learning through exploration, not a deterministic algorithm to the cars, but the true value in DeepRacer comes from the experimentation and curiosity. Also there’s loads of collateral learning – many of the racers have learned to use machine learning, reinforcement learning, python, some used Jupyter Notebooks for log analysis. Well done!
If you would like to learn with DeepRacer, head over to AWS DeepRacer page for some tips on how to get started and improve and also join the AWS DeepRacer Community where you will meet other racers and we’ll be able to learn together.