BEGIN:VCALENDAR VERSION:2.0 PRODID:-//wp-events-plugin.com//7.2.3.1//EN TZID:Asia/Kolkata X-WR-TIMEZONE:Asia/Kolkata BEGIN:VEVENT UID:22@cds.iisc.ac.in DTSTART;TZID=Asia/Kolkata:20231220T100000 DTEND;TZID=Asia/Kolkata:20231220T110000 DTSTAMP:20231211T125647Z URL:https://cds.iisc.ac.in/events/ph-d-thesis-colloquium-cds-fast-and-scal able-algorithms-for-intelligent-routing-of-autonomous-marine-vehicles/ SUMMARY:Ph.D. Thesis {Colloquium}: CDS : "Fast and Scalable Algorithms for Intelligent Routing of Autonomous Marine Vehicles." DESCRIPTION:DEPARTMENT OF COMPUTATIONAL AND DATA SCIENCES\nPh.D. Thesis Col loquium\n\n\n\nSpeaker : Mr. Rohit Chowdhury\n\nS.R. Number : 06-18-01-10- 12-18-1-16320\n\nTitle :"Fast and Scalable Algorithms for Intelligent Rout ing of Autonomous Marine Vehicles"\nResearch Supervisor: Dr. Deepak Subram ani\nDate &\; Time : December 20\, 2023 (Wednesday) at 10:00 AM\nVenue : # 102 CDS Seminar Hall\n\n\n\nABSTRACT\n\nAutonomous marine agents play a pivotal role in diverse ocean applications. These agents serve as indisp ensable instruments for acquiring crucial environmental information. They are used to explore and monitor of harsh environments\, e.g.\, to map ocea n topography\, study coral reefs\, search and rescue\, structural monitori ng of oil and gas installations etc. In naval security\, these agents are used for surveillance and strategic monitoring of maritime activities. Bui lding intelligence to optimally use these agents is essential for reducing operational costs.\n\nThe path planning problem is as follows. An autonom ous marine agent must optimally traverse from a given start location to a given target location in a stochastic dynamic velocity field like ocean cu rrents while avoiding obstacles or restricted regions in the flow. A key c hallenge is that the agent is heavily advected by the flow. The optimality may refer to minimising expected travel time or energy consumption\, data collection or risk of failure. While there are multiple methods of solvin g path planning problems\, each with its challenges\, we develop and use a fast and scalable MDP-based offline planning software that computes optim al policies\, and a novel sequence-modelling-based deep learning approach for onboard routing and dynamic planning\, where the objective is to learn optimal action sequences for the agent. The goal of this thesis is to dev elop efficient\, fast and scalable Artificial intelligence algorithms for optimal planning and on-board routing algorithms for autonomous marine age nts in stochastic dynamic environments.\n\nThe thesis comprises five works organised into two parts based on the solution approach. In the first par t\, we model the path planning problem as a Markov Decision Process (MDP) and aim to compute an optimal policy. However\, the key challenge here is that solving an MDP can be prohibitively expensive for large state and act ion spaces. To overcome this challenge\, we either approximate the optimal policy or accelerate the computation using GPUs.\n\n Physics-driven Q-le arning for onboard routing: First\, the distribution of exact time-optimal paths predicted by stochastic Dynamically Orthogonal (DO) Hamilton-Jacobi level set partial differential equations (HJLS PDEs) are utilised to lear n an initial action-value function that approximates the optimal policy. T he flow data collected by onboard sensors are utilised to get a posterior estimate of the environment. The approximated optimal policy is refined in -mission by performing epsilon greedy Q-learning in simulated posterior en vironments. We showcase the computational advantage of the approach at the cost of approximating the optimal policy.\n GPU-accelerated path plannin g: We compute an exact optimal policy by solving the path planning problem modelled as an MDP. To solve large-scale real-time problems\, which can o therwise be computationally expensive\, we introduce an efficient end-to-e nd GPU accelerated algorithm that builds the MDP model (computing transiti on probabilities and expected one-step rewards) and solves the MDP to comp ute an optimal policy. We develop methodical and algorithmic solutions to overcome the limited global memory of GPUs by using a dynamic reduced-orde r representation of the ocean flows\, leveraging the sparse nature of the state transition probability matrix and introducing a neighbouring subgrid concept to save memory and reduce the computational effort. We achieve si gnificant speedups compared to conventional sequential computation.\n Mul ti-objective GPU-accelerated path planning: The end-to-end GPU accelerated MDP solver is extended to a multi-objective path planner to solve multi-o bjective optimisation problems in path planning\, like minimising both the expected mission completion time and energy consumption. MDPs are modelle d with scalarised rewards for multiple objectives. The solver is used to s olve numerous instances of complex scenarios with other sources of uncerta inty in the environment\, enabling us to compute optimal operating curves in a fraction of the time compared to traditional solvers.\n\nIn the secon d part\, we convert the optimal path planning problem into a supervised le arning problem via sequence modelling. This approach involves learning opt imal action sequences based on the available environment information and e xpert trajectories. It also avoids the issue of re-computing optimal polic ies for onboard routing.\n Intelligent onboard routing using decision tra nsformers: We develop a novel\, deep learning method based on the decision transformer (decoder-only model) for onboard routing of autonomous marine agents. Training data is obtained from aforementioned HJLS PDE or MDP sol vers\, which is further processed to sequences of states\, actions and ret urns. The model is autoregressively trained on these sequences and then te sted in different environment settings. We demonstrate that (i) a trained agent learns to infer the surrounding flow and perform optimal onboard rou ting when the agent's state estimation is accurate\,(ii) specifying the ta rget locations (in case of multiple targets) as a part of the state enable s a trained agent to route itself to the correct destination\, and (iii) a trained agent is robust to limited noise in state transitions and is capa ble of reaching target locations in completely new flow scenarios. We exte nsively showcase end-to-end planning and onboard routing in various canoni cal and idealised ocean flow scenarios.\n Path planning with environment encoders and action decoders: We propose a novel combination of dynamicall y orthogonal flow representation with uncertainty and a transformer model (encoder-decoder) for the path planning task. We model the problem as a se quence-to-sequence translation task where the source sequence is the agent 's knowledge representation of the uncertain environmental flow. The targe t sequence is the optimal sequence of actions the agent must execute. We d emonstrate that a trained transformer model can predict near-optimal paths for unseen flow realisations and obstacle configurations in a fraction of the time required by traditional planners. Validation is performed to sho w generalisation in unseen obstacle configurations. We also analyse the pr edictions of both transformer models\, viz\, decoder only and encoder-deco der and explain the inner mechanics of learning through a novel visualisat ion of self-attention of actions and states on the trajectories.\n\n\n\n\n ALL ARE WELCOME CATEGORIES:Events,Ph.D. Thesis Colloquium END:VEVENT BEGIN:VTIMEZONE TZID:Asia/Kolkata X-LIC-LOCATION:Asia/Kolkata BEGIN:STANDARD DTSTART:20221220T100000 TZOFFSETFROM:+0530 TZOFFSETTO:+0530 TZNAME:IST END:STANDARD END:VTIMEZONE END:VCALENDAR