Learning from Human-Generated Reward (2012)
Robots and other computational agents are increasingly becoming part of our daily lives. They will need to be able to learn to perform new tasks, adapt to novel situations, and understand what is wanted by their human users, most of whom will not have programming skills. To achieve these ends, agents must learn from humans using methods of communication that are naturally accessible to everyone. This thesis presents and formalizes interactive shaping, one such teaching method, where agents learn from real-valued reward signals that are generated by a human trainer. In interactive shaping, a human trainer observes an agent behaving in a task environment and delivers feedback signals. These signals are mapped to numeric values, which are used by the agent to specify correct behavior. A solution to the problem of interactive shaping maps human reward to some objective such that maximizing that objective generally leads to the behavior that the trainer desires. Interactive shaping addresses the aforementioned needs of real-world agents. This teaching method allows human users to quickly teach agents the specific be- haviors that they desire. Further, humans can shape agents without needing pro- gramming skills or even detailed knowledge of how to perform the task themselves. In contrast, algorithms that learn autonomously from only a pre-programmed eval- uative signal often learn slowly, which is unacceptable for some real-world tasks with real-world costs. These autonomous algorithms additionally have an inflexibly defined set of optimal behaviors, changeable only through additional programming. Through interactive shaping, human users can (1) specify and teach desired behavior and (2) share task knowledge when correct behavior is already indirectly specified by an objective function. Additionally, computational agents that can be taught in- teractively by humans provide a unique opportunity to study how humans teach in a highly controlled setting, in which the computer agent’s behavior is parametrized. This thesis answers the following question. How and to what extent can agents harness the information contained in human-generated signals of reward to learn sequential decision-making tasks? The contributions of this thesis begin with an operational definition of the problem of interactive shaping. Next, I introduce the tamer framework, one solution to the problem of interactive shaping, and describe and analyze algorithmic implementations of the framework within multiple domains. This thesis also proposes and empirically examines algorithms for learning from both human reward and a pre-programmed reward function within an MDP, demonstrat- ing two techniques that consistently outperform learning from either feedback signal alone. Subsequently, the thesis shifts its focus from the agent to the trainer, describ- ing two psychological studies in which the trainer is manipulated by either changing their perceived role or by having the agent intentionally misbehave at specific times; we examine the effect of these manipulations on trainer behavior and the agent’s learned task performance. Lastly, I return to the problem of interactive shaping, for which we examine a space of mappings from human reward to objective functions, where mappings differ by how much the agent discounts reward it expects to receive in the future. Through this investigation, a deep relationship is identified between discounting, the level of positivity in human reward, and training success. Specific constraints of human reward are identified (i.e., the “positive circuits” problem), as are strategies for overcoming these constraints, pointing towards interactive shaping methods that are more effective than the already successful tamer framework.
View:
PDF
Bibtex:

Conference Presentation:
SlidesVideo
W. Bradley Knox Ph.D. Alumni bradknox [at] mit edu