A simple controller design and implementation based on Reinforcement Learning

Closed Posted Apr 7, 2014 Paid on delivery
Closed Paid on delivery

Design and implement a real target-tracking closed-loop controller for following an identified object (e.g. aircraft) as close as possible, while the object changes position in an unpredicted way:

1. In each iteration, the controller can choose between three possible actions:

“DEC”: decrease acceleration by -10%

“REM”: do not change current acceleration setting

“INC”: increase acceleration by +10%

2. Using the current distance vector, the controller can calculate the angle of the optimal acceleration vector. Alternatively, a separate set of actions can be applied for the angle itself, increasing or decreasing it accordingly (e.g. referring to standard clockwise direction), in order to bring the acceleration vector closer to the optimal trajectory.

3. The controller evaluates the “propensities” of all possible actions and selects the one that exhibits the maximum “gain”.

4. The controller receives the current distance measurement and compares it to the desired optimal distance from the central spot (= zero). The “payoff” in each iteration is inversely proportional to the difference between the current and the desired distance.

4. After such corrective action, the controller updates the “propensities” of all three possible actions using the Reinforcement Learning rules, i.e. for the action that was selected as well as for the other actions.

5. Continue the loop back from step-1, i.e. execute the next corrective action.

Engineering Java Software Architecture

Project ID: #5771858

About the project

2 proposals Remote project Active May 14, 2014

2 freelancers are bidding on average €194 for this job

trustus

Hello , We have a team of Skilled Java-J2EE professionals with experience up to 8 yrs. You will be able to directly communicate with our technical expert. Our Expertise is J2EE: 1) Frameworks: Struts More

€199 EUR in 10 days
(22 Reviews)
6.5