Citation Context

...rent approaches have been proposed. These include: (1) life-long learning, (2) imitation learning, (3) reward shaping, (4) human advice, and (5) transfer learning. Lifelong Learning Lifelong learning =-=[110]-=- was initially proposed in a supervised learning setting. This approach assumes that the learner faces a whole collection of problems over its life time. The idea is that such an agent can generalize ...

Citation Context

...ning, but serves as a perfect illustration of the ideas behind RL. 4 Chapter 1. Introduction where it does not require substantial human intervention to successfully perform transfer. As described in =-=[107]-=-, to be fully autonomous, an RL transfer agent has to perform all of the following steps successfully: • Given a target task, select an appropriate source task or set of tasks from which to transfer. ...

Citation Context

...}. The reasons behind this restriction 96 Chapter 4. Towards Automated Intertask Mappings are twofold: (1) shared information between tasks is highly probable to occur in a lower dimensional manifold =-=[113]-=-, and (2) in typical practices it is easier to define a lower dimensional space compared to a high dimensional one [9]. Algorithm 14 proceeds by projecting each of the above patterns into Sc, attainin...

Citation Context

...lications, such as [8] consider the transfer of value functions in the general game playing. Torrey et al. [112] transfers advice and transfers Q-tables through a hand coded mapping. Taylor and Stone =-=[108]-=- proposes an algorithm that is capable of transferring when the value functions between the source and target task differ in representation or the source and target reinforcement learning algorithms v...

Citation Context

...uarantees the best prediction of the dynamics of the target environment. Instance Transfer The first attempts to perform transfer between MDPs with differences in their domains are these described in =-=[104, 106]-=-. Given difference between the source and target MDPs state and action spaces, these methods make use of a hand coded intertask mapping to perform transfer. This mapping is split into two submappings....

Citation Context

...share the same actions and the human must map new sensors back into the agent space. Parameter Transfer Most of the algorithms reviewed in this section assume the presence of a hand-coded mapping. In =-=[105]-=- the notion of using a hand coded intertask mapping to transfer value functions from source to target tasks has been introduced. Its applications in [96] have shown significant improvements in complex...

Citation Context

...epresentation or the source and target reinforcement learning algorithms vary. Other methods in parameter transfer aim at automatically learning a suitable mapping between source and target tasks. In =-=[103, 104]-=- MASTER is proposed. MASTER identifies the best state mapping that guarantees the best prediction of the dynamics of the target environment. Instance Transfer The first attempts to perform transfer be...

Citation Context

...epresentation or the source and target reinforcement learning algorithms vary. Other methods in parameter transfer aim at automatically learning a suitable mapping between source and target tasks. In =-=[103, 104]-=- MASTER is proposed. MASTER identifies the best state mapping that guarantees the best prediction of the dynamics of the target environment. Instance Transfer The first attempts to perform transfer be...

Citation Context

...uch that the overall performance in a target task is improved. Human Advice Human advice aims at integrating humans to an RL agent. For instance, the human can provide action suggestions to the agent =-=[61, 111, 112]-=- or guide the agent through online feedback [43, 45]. In human advice, two major problems standout. First, there exists the need for optimally integrating the human in the loop of an RL agent, which i...

Citation Context

...on is used as an additional feature when learning in the target task. A similar method that identifies sub-tasks and features based on exploiting the source task dynamics is introduced in [24]. Walsh =-=[114]-=- deals 88 Chapter 3. Transfer for Reinforcement Learning with the more general setting in which the MDP is not discrete. The objective is to determine an aggregation of states that are capable of appr...

Citation Context

...uch that the overall performance in a target task is improved. Human Advice Human advice aims at integrating humans to an RL agent. For instance, the human can provide action suggestions to the agent =-=[61, 111, 112]-=- or guide the agent through online feedback [43, 45]. In human advice, two major problems standout. First, there exists the need for optimally integrating the human in the loop of an RL agent, which i...

Citation Context

...Matern class of covariance functions. These have the following form: kMatern = 21−ν Γ(ν) (√ 2νr l )ν Kν (√ 2νr l ) with ν, and l being positive hyperparameters and Kν being a modified Bessel function =-=[115]-=-. Please note, the scaling is chosen so that for ν → ∞ the SE covariance function is obtained. The analysis of this function in general cases, is complex and the reader is referred to [86] for a detai...

Citation Context

... advice, can also be considered as a transfer learning problem, where the knowledge is transferred from a human expert rather than other agents. In fact, such an approach has been already proposed by =-=[109]-=-. 3.2 Transfer in Reinforcement Learning Transfer learning (TL) is another paradigm created to improve learning behaviors of RL agents. The overall framework of transfer learning is shown in Figure 3....