Project $3$: Learning to Switch Traffic Lights This project is to use reinforcement learning to improve the throughput of traffic at an intersection with traffic lights. Most traffic lights operate on a predetermined phasing $($that is$,$ a timing when the lights change$),$ which does not take into account the number of vehicles that are waiting at$,$ or approaching, the intersection. This leads to a sub$-$optimal throughput of vehicles. For example, some vehicles may be waiting at a red light while the intersection is otherwise clear. A more intelligent policy, can increase the throughput. This project is to use reinforcement learning to find a policy to control the switching of traffic lights at a simulated intersection, to improve the traffic throughput. You will need to demonstrate and compare your $6$ RMIT Classification: Trusted learnt policy to a more traditional static phasing policy. REQUIREMENTS $$ This project will require you to build a simple simulation to conduct the reinforcement learning. It is advisable to build a simulation system that has configurable parameters for the flow of traffic. $$ The simulated intersection must include at least one right turning lane. An example intersection would be the intersection at Victoria Street $-$ Lygon Street. $$ Due to above work, you only need to use a single form of reinforcement learning to train an optimal policy. $$ However, you will need to also test the throughput of a static policy, and devise a suitable static policy that might be employed at a real$-$world intersection. INDEPENDENT EVALUATION $$ For your independent evaluation you must model a real$-$world intersection. $$ You will need to observe the intersection and collect basic information, such as the average throughput, and the rate at which traffic arrives at the intersection. $$ You will need to approximate the static control policy of the lights. $$ You will need to simulate both the static control policy, and learn a suitable policy for this intersection. $$ We suggest choosing a simple intersection, and observing it over a short window of around $10-15$ minutes. You may also wish to simplify the problem by excluding right$-$turn signals. $$ In case the social distancing measures do not allow you to observe a simple intersection, some statistics or a test intersection will be provided to you.