bugfix in StopAfterNoImprovement.

src/core/stop_conditions.jl

@@ -109,10 +109,11 @@ end
 # StopAfterNoImprovement
 #####
 
-Base.@kwdef struct StopAfterNoImprovement{F,B<:CircularArrayBuffer,T<:Number}
+mutable struct StopAfterNoImprovement{F,B<:CircularArrayBuffer,T<:Number}
     fn::F
     buffer::B
-    δ::T = 0.0
+    δ::T
+    peak::T
 end
 
 """
@@ -123,7 +124,7 @@ Stop training when a monitored metric has stopped improving.
 Parameters:
 
 fn: a closure, return a scalar value, which indicates the performance of the policy (the higher the better)
-e.g. 
+e.g.
 1. () -> reward(env)
 1. () -> total_reward_per_episode.reward
 
@@ -134,15 +135,15 @@ patience: Number of epochs with no improvement after which training will be stop
 Return `true` after the monitored metric has stopped improving.
 """
 function StopAfterNoImprovement(fn, patience::Int, δ::T = 0.0f0) where {T<:Number}
-    StopAfterNoImprovement(fn = fn, buffer = CircularArrayBuffer{T}(1, patience), δ = δ)
+    StopAfterNoImprovement(fn, CircularArrayBuffer{T}(1, patience), δ, T(0))
 end
 
 function (s::StopAfterNoImprovement)(agent, env)::Bool
     is_terminated(env) || return false # post episode stage
     val = s.fn()
-    improved = isfull(s.buffer) ? all(s.buffer .< (val - s.δ)) : true
+    s.peak = max(val, s.peak)
     push!(s.buffer, val)
-    !improved
+    isfull(s.buffer) ? all(s.buffer .< (s.peak - s.δ)) : false
 end
 
 #####