Projects | Academic

TensorNetworks

Sat, 27 Jun 2020 00:00:00 +0000

In this document all that is present are notes taken as the research progressed, the notes are structures in order of progression and are unfiltered raw thoughts and steps. with no respect to grammar.

We Begin by attempting to create a baseline, this is a simply CNN actor critic policy from stablebaselines 3 trained using the PPO algorithm. Once trained we transfer the parameters from pytorch to jax and buid a model, we change the policy network from a multilayerpreceptron to a TensorNetwork. (Currently progressing thus only the notes are available!)

Autonomous Driving

Mon, 27 Apr 2020 00:00:00 +0000

In this work we investigate the performance of different deep learning methods in the task of self driving, aided by regularising methods and synthetic data augmentation. We obtain results for models/methods including vanilla Multi layer preceptors, vanilla Convolutional neural networks as well as results using Transfer Learning with MobileNetV3Small and EfficentNetV2 B0 and B1, with performances ranking in the respective order from worse to best. We obtain a best performing theoretical model using the EfficentNetV2B0 model, obtaining a Kaggle MSE score of 0.01161. Fur- thermore on the practical testing, we used the MobvileNetV3Small and obtained a poorly performing model, only scoring 9/35 points in the live testing.

Electromagnetic propulsion

Wed, 27 Apr 2016 00:00:00 +0000

In this experiment, an electromagnetic propulsion system was built, with the aim to make it as reliable as possible at launching a projectile.The system consists of two solenoids and two sensors connected to an Arduino [1], the objective is to successfully accelerate the projectile through both of these solenoids without having the projectile get stuck in the middle of at the equilibrium point, therefore the purpose of the sensors is to control the flow of current, this system was tested and a success rate of 40%, 60% rate and finally 100% were achieved when all the sources of failures were formally addressed. Furthermore the experiment also looked at using a calibrated timing method, where instead of the sensors controlling the flow of current to the solenoids it was all hard coded into the python code, the optimum activation period for this design (length of each part considered) was found using trial and error, these periods were 150 milliseconds (period of the activation of the magnetic fields) from the solenoid nearest to the projectile, and 350millisecond from the other solenoid, allowing the projectile to travel from the stationary point, into the solenoid and then turns off before the projectile gets stuck in the equilibrium point (both periods start when the first signal is triggered), aside from determining the optimum times, the success rate of this method is 100%.

This experiment was carried out with the goal of making a simple electromagnetic propulsion system that’s reliable, and this was successfully achieved.

Investigating Gyroscopes

Wed, 27 Apr 2016 00:00:00 +0000

In this experiment we measured the moment of inertia of an aluminum disk using two different methods, both of which convert the system into a working pendulum, the value for the moment of inertia calculated was I = 0.0113 ± 0.002kgm2 more is displayed in the results section. Furthermore using the methods stated above, the moment of inertia of a bicycle wheel was calculated to be I = 0.14377 ± 0.016kgm2, we then used the moment of inertia and the observed period of oscillation to calculate the frequency of precession theoretically the results are displayed in the results section, on the other hand we plotted the relationship of the precession frequency and the angular frequency against the distance d (distance from the balance distance of the gyroscope), and obtained a gradient which was accurate within 5% to the true value giving us confidence in the methodology used and confidence that the theoretical value accurately described the dynamics of the gyroscopes. Finally for the nutation of the gyroscope we discover the the relationship between a nutating and a non nutating gyroscope’s period of oscillation ia a direct linear relationship, where the non nutating angular frequency is always greater that the nutating angular frequency, the percentage difference between the two was observed to always lay between -9& and -7%. More results can be found in section 4.

Investigating the Biot-Savart law

Wed, 27 Apr 2016 00:00:00 +0000

In this experiment we look at how the Biot savart law predicts the magnetic field strength in and around different current carrying objects, and different orientations, distances, currents and lengths of objects, we discover that as we keep all other variables constant and changing one variable the Biot savart law continues to predict to a fairly accurate degree how the magnetic field strength changes with that variable, with most errors calculated for out experimental results being within 6% of the calculated theoretical value, we can say that our experimental methods are indeed reliable and with certainty we can also state that the Biot savart law accurately describe the reality of how magnetic fields change relative to a certain variable for many different geometries of current carrying objects

Non-linear dimension reduction

Wed, 27 Apr 2016 00:00:00 +0000

With the increasing size and dimensionality of data, the interpretation and insight into said data becomes hard to extract. In this investigation, we look at two non-linear di- mensionality reduction methods; Laplacian Eigenmaps and Locally Linear Embeddings (LLE). We look at the intuition, mathematical formulation and effectiveness of these methods. We then apply those methods to data sampled from non-linear manifolds, and observe their resultant embedding to gain insight into how the methods behave and their effectiveness at reducing the dimensionality of the data.

We find that for some datasets LLE performs better while for others not so well, and similarly for Laplacian Eigenmaps. The datasets that we used showed that LLE was better at mapping the points to distinct values whereas, for many different choices of t, Laplacian Eigenmaps gave plots where many points overlapped. This potentially speaks to the sensitivity of the first method to the choice and proximity of neighbouring points and how it is better at preserving variation in the data. However, Laplacian Eigenmaps was able to capture an added feature of the data in that some completely distinct sections of the data were close in the 3D space.

Rutherford scattering

Wed, 27 Apr 2016 00:00:00 +0000

The purpose of this experiment is to calculate the atomic number of aluminum by knowing the depth of two different sheets and the densities of both materials then measuring the count rates of each at different angles, finally by knowing the atomic number of one of the materials we can calculate the atomic number of the other, this resulted in a calculated value for the atomic number of aluminum of 42 particles within the nucleus of an atom. Furthermore the experiment allows us to draw conclusions based on observations observed, these conclu- sions are directly responsible for the standard model of the atom and they took place as follows: • The atom contains a dense nucleus which is positively charged • The atom is mostly made of empty space • Electrons are no present in the nucleus but are in fact somewhere else. (opposing the plumb pudding model)

Structural analysis of proteins

Wed, 27 Apr 2016 00:00:00 +0000

With the goal to create a database of representations of proteins, and a further aim to help solve the protein function problem, this investigation focuses on obtaining primary protein representations derived from the protein’s 3D dimensional structure. Those representations are the 3D structure (as a base), Distogram and the Convexhull of the protein. The python classes and functions created can then be extended to include more complex protein representations, that may be used in machine learning systems, which aim at predicting a protein’s function. Currently only the Distogram can be used as input for a successful machine learning model (convolutional neural networks), other representations such as the convex hull may need more transformative steps before they can be used as inputs. Plots produced by the code are displayed inside of figures, this is to showcase the output of the code accompanying this article. Finally, no significant data generating processes have been carried out in this investigation,this is due to lack of relevance at this step and storage reasons, it is however possible to start but it wont be of any immediate significance.

The efficiency of solar cells

Wed, 27 Apr 2016 00:00:00 +0000

In this experiment we investigate how the performance of solar cells are dependant on different factors, ranging from the distance of the light source, angle of incident radiation and load to the light profile incident on the solar cells.All of these experiments resulted in a better understanding of the limitations of solar cells.

when investigating the load’s resistance we found that the power output peaks when the load resistance is equal to the internal resistance of the solar cell, this resistance was observed to be 1533Ω, with the internal resistance being measured at 1176Ω, thus giving us an error of 23.4%, this is further discussed in section 6. We also calculated the fill factor to be 0.1779.

On the other hand, when investigating the power output with respect to the angle of incidence of the incoming radiation, we found that the power increases linearly with the increasing angle of attack, this produced a maximum power output when the plane of the solar cells was perpendicular to the angle of attack and a minimum when parallel.

We also discovered that the power output obeys an inverse square relationship when the distance of the light source is increased while keeping everything else constant, leading to the conclusion that the number of photons incident is directly responsible for the generation of power. Furthermore we investigate how the light profile influences the power output, after some analysis in section 5, we managed to reach a conclusion, that being, in general terms, photons with a higher wavelength are responsibly for the majority of the power produced by the solar cells.