Contractual Reinforcement Learning:
Pulling Arms with Invisible Hands^††thanks: This work is supported in part by Army Research Office Award W911NF-23-1-0030, ONR Award N00014-23-1-2802 and NSF Award CCF-2303372.

Let us first recall the standard reinforcement learning problem in a (finite-horizon) Markov decision process $(\mathcal{A},{\mathcal{S}},\{P_{h},r_{h}\}_{h=1}^{H},P_{0})$, where we have the agent’s action space $\mathcal{A}$, the environment’s state space ${\mathcal{S}}$, the transition kernel $P_{h}:{\mathcal{S}}\times\mathcal{A}\to\Delta({\mathcal{S}})$, the expected reward function $r_{h}:{\mathcal{S}}\times\mathcal{A}\to[0,1]$, the initial state distribution $P_{0}\in\Delta({\mathcal{S}})$ and the horizon length $H$. The contractual reinforcement learning problem simply extends the MDP to a principal-agent Markov decision process $(\mathcal{A},{\mathcal{S}},\{P_{h},r_{h},c_{h}\}_{h=1}^{H},P_{0})$ with the additional cost function $c_{h}:{\mathcal{S}}\times\mathcal{A}\to[0,1]$. ¹¹1The $[0,1]$ scale of the cost and reward function range is without loss of generality, due to constant shifting and rescaling, thereby covers existing models [22, 21, 46] that assume a positive reward function for the agent. In this process, the agent interacts with the environment by taking actions and bearing the costs, whereas the principal receives the reward from the environment. Unable to directly interact with the environment, the principal has to instead design and implement contracts to incentivize the agent to take actions in her interest. Below, we formalize the design of their policies.

Following from a standard MDP, the agent’s action policy $\bm{\pi}=\{\pi_{h}:{\mathcal{S}}\to\Delta(\mathcal{A})\}_{h=1}^{H}$ specifies that at each step $h$, given the state $s$, the agent would take the action $a\sim\pi_{h}^{\bm{x}}(s)$. In the following subsection, we will discuss how the agent chooses his action policy and that it suffices to only consider deterministic action policies. Meanwhile, the principal’s contract policy $\bm{x}=\{x_{h}:{\mathcal{S}}\times{\mathcal{S}}\to\mathbb{R}_{+}\}_{h=1}^{H}$ is a sequence of non-liable payment rules $x_{h}$, where $x_{h}(s_{h},s_{h+1})$ specifies the payment to the agent if the next state $s_{h+1}$ is realized, given the current state $s_{h}$ at the $h$-th step. The non-liability constraint ensures that the principal’s payment in the contract for any realization of the next state must be non-negative; the problem would otherwise degenerate with an trivially optimal solution for the principal (see e.g., [24]). Denote $\Pi,\mathcal{X}$ as the agent and principal’s policy space, respectively. Let $|{\mathcal{S}}|=S,|\mathcal{A}|=A$ and thus $|\Pi|=(SA)^{H}$.

The typical setting of the PAMDP problems can be summarized by the following steps. In the beginning of each episode, the initial state $s_{1}\sim P_{0}$ is realized and observed by both the principal and the agent. Afterwards, the principal commits to a contract policy $\bm{x}$ and the agent accordingly chooses an action policy $\bm{\pi}$. Their interactions then proceed as follows at each step $h\in[H]$,

1. The agent takes an action $a_{h}\sim\pi_{h}(s_{h})$ and bears the cost $c_{h}(s_{h},a_{h})$. 2. The next state $s_{h+1}\sim P_{h}(s_{h},a_{h})$ is realized and observed by both the principal and agent. 3. The principal receives a noisy reward $\iota_{h}(s_{h},s_{h+1})$ and pays the agent $x_{h}(s_{h},s_{h+1})$. 4. The principal observes the agent’s action $a_{h}$.

In this step, the principal’s utility is $\iota_{h}(s_{h},s_{h+1})-x_{h}(s_{h},s_{h+1})$, her reward minus the payment to agent, whereas the agent’s utility is $x_{h}(s_{h},s_{h+1})-c_{h}(s_{h},a_{h})$, the payment from principal minus his cost. The reward noise has zero mean such that $r_{h}(s,a)=\mathop{\mathbf{E}}_{s^{\prime}\sim P_{h}(s,a)}\iota_{h}(s,s^{\prime}),\forall s\in{\mathcal{S}},a\in\mathcal{A}$. We refer the readers to Appendix B.1 for a summary of notations and Appendix B.2 for a full discussion of our modeling choices.

Without any contract design, the model reduces to a standard MDP $(\mathcal{A},{\mathcal{S}},\{P_{h},c_{h}\}_{h=1}^{H},P_{0})$ for the agent and the principal passively collects the reward from the agent’s policy. This outcome could be suboptimal for both the principal and agent. Instead, by resha** the agent’s reward environment through the design of contract policy, the principal could induce the agent adopt some action policy with higher social surplus. This motivates the problem of designing the optimal contract policy. We focus on a realistic yet challenging setup in the face of a long-lived, far-sighted and Bayesian rational agent who is also planning optimally for his cumulative reward — we expect the case of myopic agents can be worked out with simpler approach. In particular, since the agent’s utility is not necessarily $0$ under the principal’s optimal contract at any state due to moral hazard, a far-sighted agent could take certain actions that are sub-optimal in the current step, yet secure him toward certain future states where he can obtain higher cumulative utility.

We extend notions of value functions and optimal policies from MDP to PAMDP. Under any action policy $\bm{\pi}$ and contract policy $\bm{x}$, we define the principal’s state value function at the $h$-th step as,

$$V_{h}^{\bm{x},\bm{\pi}}(s):=\mathop{\mathbf{E}}\big{[}\sum_{\tau=h}^{H}r_{\tau}(s_{\tau},a_{\tau})-x_{\tau}(s_{\tau},s_{\tau+1})\big{|}\{\pi_{\tau}\}_{\tau=h}^{H},s_{\tau}=s\big{]},$$

and the agent’s state value function at the $h$-th step as,

$$U_{h}^{\bm{x},\bm{\pi}}(s):=\mathop{\mathbf{E}}\big{[}\sum_{\tau=h}^{H}x_{\tau}(s_{\tau},s_{\tau+1})-c_{\tau}(s_{\tau},a_{\tau})\big{|}\{\pi_{\tau}\}_{\tau=h}^{H},s_{\tau}=s\big{]},$$

where the expectation in both $V,U$ are with respect to the randomness of the trajectory (due to the stochasticity of state transitions and action policy). Let $V^{\bm{x},\bm{\pi}}:=\mathop{\mathbf{E}}_{s\sim P_{0}}V_{1}^{\bm{x},\bm{\pi}}(s)$ and $U^{\bm{x},\bm{\pi}}:=\mathop{\mathbf{E}}_{s\sim P_{0}}U_{1}^{\bm{x},\bm{\pi}}(s)$. The principal’s goal is to maximize her value $V^{\bm{x},\bm{\pi}}$, given the agent’s optimal response $\bm{\pi}$, which equivalently maximizes $V_{1}^{\bm{x},\bm{\pi}}(s)$ at any initial state $s$ with $P_{0}(s)>0$. Hence, we define the principal’s optimal contract policy $\bm{x}^{*}=\{x^{*}_{h}\}_{h=1}^{H}$ and the corresponding optimal value function $V^{*}$ as the optimal solution and value of the following bi-level optimization problem, ²²2Throughout this paper, we assume the agent breaks tie in favor of the principal. This is without loss of generality in generic games, since the principal can force the tie-breaking by making an infinitesimally small additional payment to the action of her interest.

$$V^{*},\bm{x}^{*}:=\mathop{maxarg}_{\bm{x}\in\mathcal{X}}V^{\bm{x},\bm{\pi}^{\bm{x}}}\quad\text{s.t.}\quad\bm{\pi}^{\bm{x}}=\mathop{argmax}_{\bm{\pi}\in\Pi}U^{\bm{x},\bm{\pi}},$$

(2.1)

where “$\mathop{maxarg}$” is a convenient operator notation on an optimization problem that returns the optimal objective value followed by its optimal solution. For notational convenience, we will denote the agent’s optimal action policy in response to contract policy $\bm{\pi}$ as $\bm{\pi}^{\bm{x}}=\mathop{argmax}_{\bm{\pi}\in\Pi}U^{\bm{x},\bm{\pi}}$, and use shorthands $V_{h}^{\bm{x}}:=V_{h}^{\bm{x},\bm{\pi}^{\bm{x}}},U_{h}^{\bm{x}}:=U_{h}^{\bm{x},\bm{\pi}^{\bm{x}}}$ for the principal’s and agent’s value function under contract policy $\bm{x}$ at the $h$-th step given that the agent responds optimally. Meanwhile, we denote $\bm{x}^{\bm{\pi}}=\mathop{argmax}_{\bm{x}\in\mathcal{X}}V^{\bm{x},\bm{\pi}}\text{ s.t. }\bm{\pi}=\mathop{argmax}_{\bm{\pi}\in\Pi}U^{\bm{x},\bm{\pi}}$ as the principal’s optimal contract policy to induce the agent’s action policy $\bm{\pi}$. We use similar shorthands $V_{h}^{\bm{\pi}}:=V_{h}^{\bm{x}^{\bm{\pi}},\bm{\pi}},U_{h}^{\bm{x}}:=U_{h}^{\bm{x}^{\bm{\pi}},\bm{\pi}}$ for the principal’s and agent’s value function under contract policy $\bm{x}^{\bm{\pi}}$ at the $h$-th step given that the agent responds optimally. Notably, since the optimization problem (2.1) hinges on the intricate correspondence between $\bm{x}$ and $\bm{\pi}$, it is unclear for now if the principal can efficiently plan his optimal policy adopting the standard approach in MDP.

To interpret the Bellman equation above, $x^{*}_{h}(s;a)$ denotes the contract with the least payment to induce the agent to take action $a$ at state $s$ in step $h$. Given that $\pi^{*}_{h}(s)$ is the best agent action for the principal to induce, the optimal contract at state $s$ in step $h$ can be determined as $x_{h}(s)=x_{h}(s;\pi^{*}_{h}(s))$. $Q_{h}^{*}(s,a),W_{h}^{*}(s,a;x)$ are respectively the principal’s and agent’s total expected utility from $h$-th step under policy $\{x^{*}_{\tau}\}_{\tau=h+1}^{H}$ and $\{\pi^{*}_{\tau}\}_{\tau=h+1}^{H}$, which can be viewed as their optimal state-action value function at $h$-th step, serving as the intermediate variable for the computation. See Appendix B.4 for the proof of correctness.

We now introduce the reinforcement learning problem in PAMDP, where the principal acts as the learner and seeks to adaptively improve its contract policy by interacting with the agent. Following the online learning convention, we use the expected regret to evaluate the learning performance in $T$ episodes, $\operatorname{Reg}(T):=\sum_{t=1}^{T}V^{*}-V^{\bm{x}^{t}},$ where $\bm{x}^{t}$ is the principal’s contract policy in the $t$-th episode.

This paper makes a few assumptions for the analysis of reinforcement learning problems. First, the far-sighted agent has perfect knowledge of his cost function and the state transition kernel $\{P_{h},c_{h}\}_{h\in[H]}$ such that he can always chooses the best response. This is realistic because in applications of our interest, agents are the experts (e.g., content creators, freelance workers, ride-sharing drivers) in the fields who has learnt about the environment sufficiently well whereas principal as the system designer does not know. Second, the agent at time $t$ is assumed to best respond to $\bm{x}^{t}$. This can be equivalently interpreted as the agent at each time $t$ showing up only once. This is motivated by the reality of Internet applications where each individual agent’s participation only accounts for a negligible portion of the system’s traffic hence has little influence over the entire system’s learning policy, so the best response (regardless of the learning policy) is optimal for each individual. Thirdly, we assume that the design space of contract is restricted to $\{x_{h}:{\mathcal{S}}\times{\mathcal{S}}\to[0,\eta]\}$ at any step $h\in[H]$. This reflects the practical concern of contract design under randomness: while contract with bounded payment may sacrifice the optimality, it regularizes the variance in the payment transfer and reduces the risk for both the principal and agents. Moreover, as long as the environment parameters have finite precision, the parameter $\eta$ can be matched to the finite bit complexity of the optimal contract. Though this assumption is without loss of generality from a modeling perspective, we expect future work to develop tighter analysis techniques to relax the dependency on $\eta$. For other regularity assumptions necessary to obtain tractable complexity results, we defer to the technical sections.

In this setup, the agent’s policy space is simply its action space $\mathcal{A}$, i.e., the set of bandit arms. $P:\mathcal{A}\to\Delta({\mathcal{S}})$ specifies an outcome distribution for each action, where the outcome space ${\mathcal{S}}$ could naturally capture the reward stochasticity of each arm in bandit learning problems. The principal designs the contract $x:{\mathcal{S}}\to\mathbb{R}_{+}$, contingent on the outcome space ${\mathcal{S}}$, to influence the agent’s choice of action. The principal’s reward and agent’s cost are both function of the agent’s action, $r,c:\mathcal{A}\to[0,1]$. We consider a contractual bandit learning problem with $T$ rounds. In the beginning of each round $t$, the principal commits to a contract $x_{t}$ and interacts with the agent as follows:

1. The agent takes the action $a_{t}$. 2. The outcome $s_{t}\sim P(a_{t})$ is realized and observed by both the principal and agent. 3. The principal receives the noisy reward $\iota(s_{t})$ and pays the agent $x_{t}(s_{t})$. 4. The principal observes the agent’s action $a_{t}$.

Here, the noisy reward function satisfies $\mathop{\mathbf{E}}_{s\sim P(a_{t})}\iota(s)=r(a_{t})$, and we assume the agent’s action always maximizes his expected utility, i.e., $a_{t}\in\mathop{argmax}_{a\in\mathcal{A}}\{P(a)\cdot x_{t}-c(a)\}$. To determine the principal’s optimal contract, let us recall the notion of least payment function from the general setup. We similarly define $\zeta:\mathcal{A}\to\mathbb{R}_{+}$ such that for any given action $a\in\mathcal{A}$, it outputs the least amount of the expected payment necessary to induce $a$, $\zeta(a):=\mathop{min}_{x\in\mathcal{X}^{a}}P(a)\cdot x,$ where $\mathcal{X}^{a}=\{x\in\mathcal{X}:[P(a)-P(a^{\prime})]\cdot x\geq c(a)-c(a^{\prime}),\forall a^{\prime}\neq a\}$ denotes the set of all contracts under which the agent would respond with action $a$. Hence, the principal can determine the optimal action to induce, $a^{*}=\mathop{max}_{a\in\mathcal{A}}\sum_{t=1}^{T}[r_{t}(a)-\zeta(a)]$ with the optimal contract $x^{*}=\mathop{argmin}_{x\in\mathcal{X}^{a^{*}}}P(a^{*})\cdot x$. With the benchmark of the optimal contract $x^{*}$ that induces $a^{*}$ with the least payment $\zeta(a^{*})$, we can measure the learning performance in $T$ rounds with the expected regret as follows, $\operatorname{Reg}(T)=\mathop{max}_{a^{*}\in\mathcal{A}}\sum_{t=1}^{T}\big{[}r(a^{*})-\zeta(a^{*})\big{]}-\sum_{t=1}^{T}\big{[}r(a_{t})-P(a_{t})\cdot x_{t}\big{]}.$ This problem is a strict generalization of standard online learning, as it degenerates to the standard notion of regret when $\zeta(a)=0,\forall a\in\mathcal{A}$. However, with the additional $\zeta$ function, the no-regret learner must not only obtain good estimation of both $r$ and $\zeta$ towards the optimal action, but also implement the contracts that induce the optimal action and have expected payment approaching towards $\zeta$.

We begin with a natural assumption that enable us to simply employ existing techniques in online learning to obtain tractable complexity results for a large class of contractual bandit learning problems.

This assumption ensures the regularity of the problem instance in the sense that each action is dominantly capable of inducing a set of outcomes over others such that for any cost function $c$ and any action $a$, there exists a contract $x$ to induce $a$. To see this, one can explicitly construct such contract as $x=e\mathop{max}_{a^{\prime}}\frac{c(a)-c(a^{\prime})}{\lambda}$, where $e$ is the event such that $[P(a)-P(a^{\prime})]\cdot e\geq\lambda,\forall a^{\prime}\neq a$. Otherwise, if $\lambda\leq 0$, then there could be some action that is never the agent’s best response under any contract.

We now propose a generic approach to design statistically efficient algorithm for contractual bandit learning problem. The key idea of our approach is to decouple the learning of the contract from the learning of the optimal action. In particular, let us first assume an oracle in Definition 1 that is able to construct a robust contract set for each action $a\in\mathcal{A}$, despite the uncertainty in parameter estimation. We use the robust contract set to determine the optimistic action and eventually learn the optimal action with no regret. This enables us to decouple the sample complexity result into the estimation errors from optimal contract and the optimal action, according to Theorem 2.