Planning with Uncertainty Sample Clauses

Planning with Uncertainty. ‌ Working with robotic agents leads to planning solutions needing to address the issues that appear when having uncertainty in the planning or execution phase. Uncertainty can appear either due to imperfect reading of sensors resulting into the agent being uncertain of its state or uncertainty in the execution due to actions not failing. Solutions to this have varied from using different types of planning types (i.e. probabilistic planning, contingent planning), to approaches based on machine learning techniques. Planning in robotic agents took two main approaches, one by ignoring the action uncertainty and trying to deal with how to act when it will fail, and the other approach was by incorporating the uncertainty in the planning model. One approach on reasoning about an agent’s execution when there is a high chance the plan might fail, is by preparing the system for failure and reasoning about replanning routines. This approach was first described by Xxxxx in [59], who proposed reusing the data structures (subgoal trees and decision graphs) that were used for the initial plan to guide replanning. More recent work on replanning, Xxxx et al. created FF-Replan [146], which is a planner that takes as input a probabilistic model, and then determinises the problem into a classical planning problem. It will find a plan for the problem by using the FF-planner [69], and if a failure happens during execution, it will plan again but in the same determinisation of the problem. The solution presented by Xxxxxx et al. [10] was to extend ROSPlan to accept any prob- abilistic planners based on RDDL models [126]. This way, uncertainty could be taken into account from the beginning, and instead of giving a plan as a solution, it would return a policy (mapping of states to actions). Machine learning techniques have also been used, as done by Krivic et al. [89], who tackled the issue of uncertainty in the state the agent is in. Their solution was to approximate the values of the unknown planning variables by using Machine Learning classifiers trained on previous solutions. BDI approaches differ from planning with uncertainty, as the agents are not expected to create neither a plan, nor a policy, but to select the appropriate Intention Plan (I-Plan) depending on the desires of the agent and the state of the environment. Our work focused on the idea of using I-Plans and the creation of plan libraries, for example using implementations such as PRS [77, 121] or Xxxxx [7]. While...